In March 2024, the Securities and Exchange Commission fined two investment advisers, Delphia (USA) Inc. and Global Predictions Inc., a combined 310,000 dollars for telling investors their artificial intelligence did more than it actually did. Two years later, that enforcement logic has moved from the asset management industry into the venture capital term sheet. A founder who cannot document what a third-party large language model vendor does with customer data, which state privacy statutes apply to that data, and what the board actually reviewed before relying on artificial intelligence now faces the same category of scrutiny a public company faces from the Commission’s Enforcement Division.

The mechanism is simple once traced to its source. Section 17(a) of the Securities Act of 1933 and Rule 10b-5 under the Securities Exchange Act of 1934 make a false statement about a product’s artificial intelligence fraud, whether the security is a public share or a Regulation D private placement sold under Rule 506(b) or 506(c). The Commission’s Enforcement Division built a Cyber and Emerging Technologies Unit in February 2025 specifically to read artificial intelligence disclosures inside Forms 10-K, 10-Q, and 8-K, and the North American Securities Administrators Association named artificial intelligence marketing claims a 2026 examination priority for state registered advisers. None of that authority is limited to companies that have already gone public. It reaches the representations a founder makes in a subscription agreement the same day the wire arrives.

How does a third-party large language model vendor create diligence risk

The answer sits in two documents most founders have never read closely: the vendor’s Data Processing Addendum and its training exclusion policy. OpenAI does not use API or Enterprise data to train its models by default. Anthropic reduced API log retention from thirty days to seven days as of September 14, 2025, and deletes inputs and outputs automatically unless a customer negotiates a longer window through its own Data Processing Addendum. Both vendors offer a zero data retention agreement to qualifying enterprise customers, removing stored inputs beyond what abuse screening requires. A startup that cannot produce a signed Data Processing Addendum, a documented training exclusion election, and a written record of its retention window is handing a diligence team an open item, and diligence teams in 2026 are trained to find it. The Federal Trade Commission’s settlement with DoNotPay, a 193,000 dollar final order approved in February 2025 over an AI product marketed as performing to the standard of a licensed attorney without substantiating testing, and its separate 18 million dollar stipulated judgment against Air AI, show what happens when a company’s public claims about its artificial intelligence outrun what it can prove. Both actions rested on Section 5 of the Federal Trade Commission Act, the same statute that reaches a startup’s own claims about its vendor relationship.

Which state privacy statutes actually apply to a startup’s customer base

Twenty states had comprehensive consumer privacy statutes in effect at some point during 2026. Indiana, Kentucky, and Rhode Island took effect January 1, 2026, and Connecticut, Arkansas, and Utah added amendments effective July 1, 2026. California added its own layer on top of the existing California Consumer Privacy Act: the Privacy Protection Agency finalized rules on automated decisionmaking technology, risk assessments, and cybersecurity audits effective January 1, 2026, phased by revenue through April 1, 2030 for the smallest covered businesses. Colorado’s approach to artificial intelligence specifically has changed twice since 2024. Senate Bill 24-205 was set to take effect February 1, 2026. Colorado’s governor signed Senate Bill 25B-004 in August 2025, moving that date to June 30, 2026, then signed Senate Bill 26-189 in May 2026, replacing the original statute with a new artificial intelligence law effective January 1, 2027. A data map naming every applicable state, the threshold that triggers coverage, and a completed risk assessment for any automated decisionmaking use is now standard diligence, not an artifact investors expect a company to build after signing.

What board records satisfy Delaware’s oversight duty for artificial intelligence

Delaware’s Court of Chancery, applying the duty recognized in In re Caremark International Inc. Derivative Litigation, extended director oversight obligations to cybersecurity risk in 2025 for any company that stores consumer data or depends on digital infrastructure. A director does not need to understand how a transformer model produces an output. The board needs to show it received regular briefings on how the company relies on artificial intelligence and what could go wrong. On December 4, 2025, the Commission’s own Investor Advisory Committee recommended that companies disclose how they define artificial intelligence, what oversight mechanism the board uses, and what material effects that use has produced. That recommendation targets public companies, but it is already shaping what venture investors expect to see in board minutes at private companies preparing to raise. A board package that shows the company mapped its uses of artificial intelligence, measured the associated risk, assigned management responsibility, and kept the board briefed, consistent with the four functions of the National Institute of Standards and Technology’s AI Risk Management Framework, gives investor’s counsel something concrete if a Caremark claim ever follows the round.

The founders who close rounds fastest in 2026 are not necessarily the ones with the strongest product demonstration. They are the ones who can hand a data room three things without delay: an executed vendor Data Processing Addendum, a state by state privacy map, and board minutes that show artificial intelligence reliance was actually discussed before the company relied on it.

Read my full analysis here: https://theinnovationattorney.com/blog/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Institutional investors are treating a written founder vesting schedule with a one year cliff, tracking NVCA Investor Rights Agreement Section 5.3, as a condition to closing a priced round in 2026. A lead investor’s term sheet increasingly lists a signed restricted stock purchase agreement among the conditions to closing rather than as a side letter to negotiate after the wire lands. Founders who never documented vesting, or whose shares fully vested years before this round, are the ones asking what specific reverse vesting and company buyback language a term sheet will require before a lead investor signs.

The question surfaces at an inconvenient moment. By the time a term sheet reaches signature, the company has already negotiated valuation, board composition, and protective provisions. Vesting shows up in the conditions to closing section almost as an afterthought, yet it is often the single item standing between a signed term sheet and a closed round, because the lead investor’s counsel will not release funds until every founder has executed the restricted stock purchase agreement the term sheet references.

How the standard schedule works

NVCA Investor Rights Agreement Section 5.3 supplies the template most lead investors hand to portfolio companies: vesting over forty eight months, with nothing vesting during the first twelve. Twenty five percent vests on the cliff date, and the remaining seventy five percent vests monthly over the following three years. When a founder incorporated without documenting a vesting schedule, this is the schedule the round will impose retroactively, dated to a vesting commencement date the parties negotiate, often the company’s original incorporation date or the date the founder began working full time.

What the repurchase right actually says

Reverse vesting issues a founder’s full share grant up front and then gives the company a contractual right, not an automatic forfeiture, to repurchase the unvested portion if the founder departs before the shares vest. Delaware General Corporation Law Section 160 supplies the corporate authority for that repurchase, subject to the rule that a company cannot buy back its own stock if doing so impairs capital. The 2023 amendments to Section 160 clarified how the resulting treasury shares may later be resold under Section 153, a detail that matters when a company wants to reissue repurchased founder shares to a replacement executive rather than retire them. The restricted stock purchase agreement itself fixes the repurchase price, commonly the lower of original issue price or fair market value, and a repurchase window, commonly ninety days from termination, during which the board must act or the option lapses.

Does the cliff apply to shares already earned

This is the specific ambiguity founders are running into. Where a founder already vested shares under an earlier grant and is only placing new or unvested shares onto a schedule at the priced round, a fresh one year cliff is hard to justify on the merits: the founder has already demonstrated the staying power a cliff exists to test. SaaStr founder Jason Lemkin has taken exactly this position in practitioner commentary, arguing that reverse vesting at a financing generally should not carry a new cliff for founders who are already committed. Some lead investors propose the NVCA cliff language anyway, as a matter of drafting convenience rather than a considered position on a founder who already has three years of history behind the company. That gap between boilerplate and rationale is precisely where a founder’s counsel should redline.

Credit for time served, and what happens without it

When a term sheet requires revesting, it typically places some or all of a founder’s previously vested shares back onto a new schedule as of closing. Founders routinely negotiate credit for time already served, so that only a portion, often twenty five to fifty percent of total founder shares, goes back onto the new schedule rather than the full grant. None of this happens automatically. A term sheet silent on credit for time served will be read by the lead investor’s counsel as requiring full revesting from zero, because silence favors the drafting party’s default position, not the founder’s expectation.

The tax deadline that follows the signature

A founder who signs a new restricted stock purchase agreement covering shares subject to company repurchase faces a substantial risk of forfeiture under Internal Revenue Code Section 83(c), and with it a firm thirty day deadline under Section 83(b) to elect to be taxed at grant rather than as the shares vest. The Internal Revenue Service introduced Form 15620 in November 2024 as the first official form for this election, and added an online filing option in 2025. Courts have consistently declined to excuse a late filing under Section 83(b), so the thirty day clock needs to be calendared from the date the new restricted stock purchase agreement is signed, not from the date the round closes, a distinction that has cost founders their election in past financings.

The counterweight founders negotiate in return

Founders typically negotiate double trigger acceleration alongside reverse vesting: unvested shares accelerate, often in full, only if the company experiences a change of control and the founder is then terminated without cause or resigns for good reason within a defined window, commonly twelve months, after that change of control. Roughly eighty five to ninety percent of Series A and later stage term sheets in 2026 include some form of double trigger acceleration for the founding chief executive. A founder signing a fresh reverse vesting schedule at closing without also negotiating double trigger acceleration is accepting the restriction without the protection that has become standard alongside it.

In three decades of counseling companies through priced rounds, the redline that most often gets missed is not the vesting percentage. It is the silence on credit for time served and the absence of an explicit carve out for the one year cliff on shares a founder already earned. Those two gaps cost founders equity quietly, long after the term sheet is signed and forgotten.

Read my full analysis here: https://theinnovationattorney.com/unlocking-venture-financing-founder-vesting-buybacks-and-tax-compliance/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Tennessee, Utah, Oklahoma, Louisiana, and Idaho have each signed a non binding agreement with the Department of Energy to host a Nuclear Lifecycle Innovation Campus, a facility combining uranium enrichment, fuel fabrication, spent fuel research, and waste storage on one site. The Department expects to narrow the five contenders to three finalists before the end of 2026. Utilities that have spent nearly three decades collecting Judgment Fund damages for spent fuel the federal government was contractually obligated to remove starting in 1998 are watching closely, because the plan asks Congress to grant authority two federal appeals court rulings already said the Nuclear Regulatory Commission does not have.

The spent fuel problem is older than most of the reactors producing it today. Congress passed the Nuclear Waste Policy Act in 1982, funded by a fee of one mill per kilowatt hour on every unit of nuclear electricity generated, and directed the Department of Energy to open a permanent repository by 1998. In 1987, Congress narrowed the search to a single site, Yucca Mountain, Nevada, and in 2002 it overrode Nevada’s formal objection to complete the designation. The site never opened. Funding for its licensing process ended in 2011, and roughly one hundred thousand metric tons of spent commercial fuel now sit in pools and dry casks at more than seventy reactor locations across the country, growing by about two thousand metric tons a year.

Distributing the function across five willing states is a reasonable departure from four decades of trying to force a single site through litigation. Oak Ridge, Tennessee, and Idaho National Laboratory each carry existing federal nuclear infrastructure and a trained workforce that a green field site would lack, and the Department’s new Center for Spent Fuel Research at Idaho National Laboratory, announced in January 2026, gives the plan a research foundation rather than a blank page. The remaining obstacle is not geology or willingness. It is statutory authority. In 2024, the Fifth Circuit Court of Appeals vacated Nuclear Regulatory Commission licenses for consolidated interim storage facilities in Texas and in Lea County, New Mexico, ruling that neither the Atomic Energy Act nor the Nuclear Waste Policy Act gives the Commission power to license privately owned storage away from a reactor or a federal repository. Holtec International, the New Mexico licensee, withdrew its project in 2025 after the Supreme Court resolved a related challenge on procedural grounds without reaching that statutory question. Every state agreement signed this year is non binding until Congress supplies the authority those courts said the agency lacks.

Three environmental risks apply to consolidated storage: transportation accidents, cask or container failure, and handling errors, and most already have a demonstrated safeguard elsewhere in the federal system. Transport casks must survive a thirty foot drop, a thirty minute fire, and submersion under Nuclear Regulatory Commission rules at 10 CFR Part 71, and the agency’s own risk studies rate the probability of a radiological release during shipment as very low. Storage has one clear cautionary tale. On February 14, 2014, a drum of transuranic waste at the Waste Isolation Pilot Plant near Carlsbad, New Mexico, ruptured after technicians packed it with an organic absorbent instead of the mineral based material its chemical profile required, producing a reaction that blew off the lid and contaminated about thirty five percent of the underground disposal area. The facility closed for nearly three years, and direct cleanup ran to roughly five hundred million dollars. Investigators traced the rupture to a packaging procedure rather than to any flaw in the salt formation, which performed as designed, and it produced a permanent fix: the Environmental Protection Agency and the New Mexico Environment Department now require overhauled waste characterization at every generator site before a single drum ships. Any new campus inherits that same regulatory model, plus dry cask designs the NRC projects will remain safe for more than a century.

The reprocessing question has a simpler answer than most people expect. The United States does not recycle spent fuel commercially because of a decision made in 1977, not because the chemistry is impossible. That year, following a Ford Foundation study warning that separated plutonium from commercial reprocessing could hasten weapons proliferation, the federal government deferred reprocessing indefinitely and canceled the Clinch River Breeder Reactor program. A later administration lifted the formal ban in 1981, but by then a 2003 Harvard Belfer Center analysis had already shown what utilities suspected: reprocessing only becomes cheaper than direct disposal once uranium prices exceed roughly three hundred sixty dollars per kilogram, a level the market has rarely reached. France chose differently and has recycled fuel at La Hague since 1966, recovering plutonium that its Melox plant turns into mixed oxide fuel supplying about ten percent of French nuclear electricity. Reprocessing still leaves high level waste that needs its own repository, a French project called Cigeo, so the French approach shrinks the disposal problem without solving it entirely. Current federal research favors pyroprocessing and electrorefining over the plutonium separating methods France uses, because those techniques keep plutonium mixed with other actinides rather than isolating it. A 2026 study led by Oklo with Argonne and Idaho National Laboratory found that waste from an electrorefining process is compatible with deep borehole disposal, placing waste as far as five kilometers underground, an option still years from commercial use but no longer purely theoretical.

The five state plan will succeed or fail on a single fact that has nothing to do with geology, chemistry, or engineering: whether Congress passes the authority two federal appeals courts have already said the executive branch does not have on its own. Utilities holding more than thirty four billion dollars in accumulated damages claims have every reason to push that legislation forward, and investors pricing new reactor construction into any of these five states should be pricing that legislative outcome first.

Read my full analysis here: https://theinnovationattorney.com/blog/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Federal prosecutors in the Northern District of Georgia charged Samuel Tunick under 18 U.S.C. Section 2232(a) in November 2025, the first known application of that destruction of property statute to a phone’s self triggered data wipe. Tunick had entered a duress PIN built into GrapheneOS while Customs and Border Protection agents held his phone at Hartsfield-Jackson Atlanta International Airport, and the device erased its contents instead of opening. He faces up to five years in prison for using a security feature he did not build, under a statute written for smugglers who dumped contraband overboard.

The stop happened on January 24, 2025, as Tunick returned from a trip to the Dominican Republic. Agents pulled him into secondary inspection, told him a warrant was not required, and asked him to open his phone. Court filings describe what happened next. The screen went blank, flashed several times, and the phone appeared to restart. Agents seized the device anyway and released Tunick a short time later. His attorneys say the agents cited a search for child exploitation material without evidence to support it, and that the real interest was Tunick’s association with Defend the Atlanta Forest, the movement opposing the Atlanta Public Safety Training Center.

What is a duress PIN and how does GrapheneOS implement it

A duress PIN is a second passcode that does not open the phone. Entering it instead destroys the cryptographic keys protecting the phone’s encrypted data, an action indistinguishable from the outside from an ordinary restart. GrapheneOS built the feature without a confirmation prompt or any visible warning, so a person entering the code under pressure gives no sign of what the code actually does. In drafting incident response policies for hardware clients, the recurring gap is not the encryption itself. It is the absence of a plan for what happens when a device is seized, a gap this case now shows can carry criminal exposure rather than only a data loss.

Does the border search exception let agents search a phone without suspicion

In the Eleventh Circuit, yes. The border search exception traces to United States v. Ramsey, 431 U.S. 606 (1977), and gives the government authority to search people and property at the border without a warrant or individualized suspicion. Circuits disagree on how far that authority extends to phones. The Ninth Circuit required reasonable suspicion for a forensic search in United States v. Cotterman, calling it a computer strip search. The Fourth Circuit required individualized suspicion of a transnational offense in United States v. Kolsuz. The Eleventh Circuit rejected both positions in United States v. Touset, holding that a forensic search of a phone needs no more suspicion than a search of a suitcase. Because Tunick’s case sits in the Northern District of Georgia, Touset controls. In July 2026, the Fourth Circuit added United States v. Belmonte Cardozo, holding that even a manual phone search is routine and needs no suspicion, while the Electronic Frontier Foundation and the National Association of Criminal Defense Lawyers keep pressing appellate courts for a probable cause and warrant standard.

Can the government force you to reveal your phone passcode

No. The Fifth Amendment privilege against self incrimination protects a traveler from being forced to disclose a memorized passcode, under the act of production doctrine from Fisher v. United States, 425 U.S. 391 (1976). The Eleventh Circuit applied that doctrine to encrypted devices directly in a 2012 ruling, In re Grand Jury Subpoena Duces Tecum, finding that compelled decryption can violate the privilege unless the government already knows the person can access the device. Tunick was never asked to explain a key. He entered a code. That single fact is where the case’s novelty begins. Agents got a code, and the code they got was engineered to protect data rather than expose it.

Is a data wipe destruction of property under Section 2232(a)

That is the statutory question the court still has to answer. Section 2232(a) punishes destroying, damaging, wasting, disposing of, or otherwise taking action against property to prevent the government from taking lawful custody of it. Every prior prosecution under the statute involved a physical object: narcotics thrown from a boat, a vessel scuttled to avoid forfeiture, paper records burned during a raid. Tunick’s phone was never damaged. The government has physical custody of it right now. Defense counsel will argue that data is not the kind of property Section 2232(a) contemplates. Prosecutors will argue that the statute’s language covers any action taken to defeat the government’s authority over the object of a search, regardless of whether the object is physical or digital.

Why do Tunick’s attorneys call this a selective prosecution

Because the stated reason for the search does not match, in their telling, the reason agents actually cared about Tunick. Under United States v. Armstrong, 517 U.S. 456 (1996), a defendant claiming selective prosecution must show that similarly situated people were treated differently and that the decision to search or charge him was driven by an improper purpose. Tunick’s team argues the improper purpose here was his connection to Defend the Atlanta Forest. A ruling on the suppression motion is expected later in 2026.

The ruling will decide more than one case. It will tell millions of travelers who run privacy hardened phones whether that choice creates criminal exposure the moment they cross a border. It will also test whether a decades old statute, written for smugglers and physical contraband, extends to a line of code that erases a key.

Read my full analysis here: https://theinnovationattorney.com/wiping-data-at-the-border-5th-amendment-limits-and-criminal-exposure/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

On February 27, 2026, the Delaware Supreme Court closed a fourteen month challenge to Senate Bill 21 and removed the last legal obstacle to a statute that Tesla chief executive Elon Musk had pushed for since the Court of Chancery voided his $56 billion pay package in early 2024. In Rutledge v. Clearway Energy Group LLC, the court held that amended Section 144 of the Delaware General Corporation Law is constitutional, that it does not strip the Court of Chancery of its authority to hear fiduciary duty claims, and that its retroactive application to transactions predating March 2025 does not violate due process. For a board that includes a director, officer, or stockholder with an actual conflict, the ruling answers the question that matters in practice: which approval process now keeps a related party transaction out of an entire fairness trial.

Section 144 has always governed transactions where a director or officer stands on both sides of a deal. Senate Bill 21, signed into law in March 2025 after passing the General Assembly by the supermajority required for changes to Delaware’s corporate code, extended that statute to controlling stockholders and added a bright line definition of control: a person who owns a majority of a corporation’s voting stock, or a person who owns at least one third (33.3%) of the voting stock and also exercises actual managerial authority over the company. Before the amendment, Delaware courts decided control case by case, which meant a large minority investor with board seats and a veto right could be treated as a controller years after a transaction closed, with the harsher entire fairness standard applied retroactively to the deal.

How Does The Committee Approval Safe Harbor Work

A related party transaction, other than a going private transaction, is shielded from an injunction and from a damages award if it is approved or recommended in good faith by a committee of the board made up of at least two disinterested directors. The committee has to be informed, meaning it has the time, the advisors, and the facts needed to evaluate the deal on its own terms, and it has to act with the same care a disinterested board would use for any material transaction. Rutledge did not supply a checklist for what informed enough looks like, and the Court of Chancery kept jurisdiction to decide that question on a transaction by transaction basis. What changed is the consequence of getting it right. Once a qualifying committee approves the deal, review shifts from the price the controller paid to the process the committee ran, and the business judgment rule, not entire fairness, becomes the standard a plaintiff has to overcome.

How Does The Stockholder Ratification Safe Harbor Work

The second path runs through the stockholders instead of the board. A related party transaction also earns safe harbor treatment if a majority of the shares held by disinterested, non controlling stockholders approve it on a fully informed and uncoerced basis. The proxy statement or information statement supporting that vote has to disclose the controller’s interest and the material terms of the deal in enough detail that a reasonable stockholder can evaluate it, since a vote obtained through incomplete disclosure does not count as informed under the statute. A board can rely on either the committee path or the stockholder path for an ordinary related party transaction. It does not need both.

Why Do Going Private Deals Still Need Both Protections

Going private transactions did not get the same discount. Section 144(c) preserves the dual protection standard the Delaware Supreme Court built in Kahn v. M & F Worldwide Corp. in 2014: business judgment review is available only if the controller conditions the deal, from the first offer, on approval by both an independent special committee given the authority to negotiate price and to reject the transaction outright, and an informed, uncoerced vote of the disinterested minority. Miss either condition, and the deal reverts to entire fairness with the controller carrying the burden of proving a fair price and a fair process. The General Assembly left this standard alone because a going private deal, unlike an ordinary related party transaction, ends the minority’s ability to exit by selling the stock, which calls for a stricter check on the controller’s conduct.

What Did The Delaware Supreme Court Decide In Rutledge V. Clearway Energy Group

The case arose out of a derivative suit brought by a Clearway Energy, Inc. stockholder against its majority stockholder, Clearway Energy Group LLC, and the company’s former chief executive, alleging that the board overpaid the controller for an interest in an Idaho wind project. The plaintiff asked the Court of Chancery to declare Section 144 unconstitutional on two grounds: that it stripped the court of its constitutionally protected equity jurisdiction, and that its retroactive application took away a vested property right in his claim without due process. The Supreme Court rejected both arguments in a unanimous opinion. On jurisdiction, the court held that Section 144 changes the standard of review and narrows the available remedies, but it leaves every fiduciary duty claim, including whether the statutory prerequisites were satisfied, with the Court of Chancery. On retroactivity, the court held that a plaintiff has no vested right in a particular common law standard of review, and that the General Assembly’s stated interest in giving Delaware corporate law predictability satisfies due process, even for claims tied to conduct that predates the March 2025 enactment.

In redrafting committee charters for controller transactions since the ruling, the language that draws the most scrutiny is not the fairness opinion. It is the sentence describing whether the committee could actually say no, because that is the sentence Rutledge told the Court of Chancery to test first.

Rutledge settles the constitutional question, but it does not settle how much disclosure counts as informed, or how a company with an investor sitting at exactly one third ownership proves that investor lacks actual managerial authority.

Read my full analysis here: https://theinnovationattorney.com/blog/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

A most favored nation clause in a Simple Agreement for Future Equity can retroactively reprice an early, cheap SAFE down to match a lower cap issued months later in the same bridge, without a dollar of new cash entering the company. Most favored nation provisions now appear in 35 to 40 percent of pre-seed angel rounds, shifting a founder’s effective dilution by 15 to 40 percent depending on the gap between the earliest cap and the lowest cap in the bridge. The valuation cap on the SAFE itself never touches the protective provisions, board seats, or participation rights that a priced round investor will negotiate separately, months after the SAFE is signed. Those terms live in the certificate of incorporation and the voting agreement, not the cap.

Seed investors are once again requiring protective provisions and consent rights that were commonly waived during the 2021 financing peak, when capital was abundant and founders held most of the negotiating position. The pendulum swung back once deal volume cooled: protective provisions, the list of company actions requiring investor consent before a company can issue new equity, take on debt above a stated threshold, or sell itself, now appear in more than 90 percent of seed and Series A deals. Liquidation preference terms softened even as consent rights hardened. Ninety eight percent of United States venture rounds in the second quarter of 2025 used a one times non-participating preference, the structure where an investor recovers its investment first and then only shares in additional proceeds by converting to common stock. The participation economics improved for founders while the control mechanics reverted to their pre-2021 form, and a founder who raised on a SAFE during the intervening years signed an instrument that says nothing about either shift.

A most favored nation clause gives an early SAFE holder the right to swap into the terms of any subsequent SAFE the company issues on better terms, typically a lower valuation cap or a larger discount rate, until the clause expires or the company closes its priced round. One version of the post-money SAFE format is built entirely around this mechanic: an instrument with neither a cap nor a discount, its terms fixed only by whatever the next SAFE investor negotiates. A founder who signs a 10 million dollar cap SAFE in January and an 8 million dollar cap SAFE in June has, absent an explicit carve out in the January instrument, just repriced the January investor down to 8 million dollars. The company received no additional benefit from that repricing; the January investor simply receives more shares at conversion for the same cash already in the bank.

Four specific drafting choices preserve a founder’s position against both risks. Scope the most favored nation clause to price terms only, meaning the valuation cap and the discount rate, and expressly exclude side letter rights such as pro rata rights, information rights, or a board observer seat granted to a later investor from the comparison. A most favored nation clause written broadly enough to sweep in those side letter terms hands the earliest, often smallest, check a claim on decision rights that investor never bargained for directly. Attach an expiration date to the clause, tied to the earlier of a stated date or the closing of the priced round, so the right does not survive indefinitely into financings the original investor had no part in shaping. Where a bridge requires more than one SAFE, use a single valuation cap across every instrument in that bridge rather than raising sequential SAFEs at different caps; a uniform cap removes the retroactive repricing risk because there is no gap for a most favored nation clause to close.

The valuation cap protects a founder from only one kind of surprise: a runaway price. It does nothing to protect against the certificate of incorporation, voting agreement, and investor rights agreement that govern the preferred stock the SAFE eventually converts into. Delaware General Corporation Law Section 242(b)(2) already gives holders of a class of stock a vote on any charter amendment that adversely affects that class, so a protective provision restating that right is enforcing a statutory floor rather than creating a new right. A founder negotiating the priced round should push for a single class vote threshold, requiring approval from a majority of the preferred stock voting together, rather than a separate veto for each series or each individual investor. A protective provision that lets any single investor block a financing, a sale, or a debt incurrence regardless of that investor’s check size is a term that outlives the SAFE and governs the company for years.

In my own practice, the moment this becomes visible to a founder is rarely at signing. It is eight months later, when a priced round term sheet arrives and the cap table software flags a conversion number nobody in the room expected, because an early SAFE’s most favored nation clause quietly matched a cap negotiated with a completely different investor in a completely different month.

Delaware case law leaves open whether a protective provision drafted only to track Section 242(b)(2) reaches a merger in which the surviving corporation’s certificate is amended, absent express language extending that right to mergers. A founder who assumes seed stage protective provisions already cover that scenario should confirm the certificate says so directly. Model SAFE and side letter language may also be revised across the industry to cabin most favored nation scope to price terms alone, closing the gap that currently lets the clause sweep in decision rights granted to a later investor.

Read my full analysis here: https://theinnovationattorney.com/blog/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Connecticut Senate Bill 1295, Arkansas House Bill 1717, and Utah House Bill 418 take effect July 1, 2026, each imposing a distinct data minimization and profiling disclosure standard. A single data protection assessment built to Connecticut’s new profiling impact assessment content list will not by itself satisfy Arkansas’s minor specific minimization duty, and it will be surplus paperwork under Utah’s narrower amendment, which adds only a right to correct. A software as a service company collecting consumer data in all three states needs one minimization narrative supported by three state specific annexes, not one uniform document.

The three amendments arrived from the same 2025 legislative cycle and landed on the same effective date by coincidence of session timing rather than coordination. That timing has convinced a number of founders that the substance converged too. It did not. Connecticut rewrote the general consumer minimization standard and built a profiling assessment with defined content requirements. Arkansas left its general consumer statute alone and instead built a second, narrower law for a different population. Utah added a single consumer right and touched neither minimization nor profiling. Reading all three as parallel updates to a shared floor is the mistake this analysis exists to correct.

What Connecticut actually changed.

Connecticut’s Public Act No. 25-113 replaces the requirement that collection be adequate, relevant, and reasonably necessary with a requirement that it be reasonably necessary and proportionate to the disclosed purpose. Proportionate is the word doing the work: a controller can no longer justify a data field on relevance alone, the sensitivity and volume collected now has to match the weight of the purpose. The amendment also strips the word solely from the profiling opt out right, so a consumer can object to profiling behind any automated decision with a legal or similarly significant effect, including one made on a controller’s behalf by a processor or vendor. Where that threshold is met, the controller must complete an impact assessment covering seven specific elements: purpose and intended use, a risk and mitigation analysis, the data categories used and outputs generated, the categories used to tailor the result, performance metrics and known limitations, consumer notice during the profiling, and post deployment oversight. The assessment duty reaches only profiling activities created or generated on or after August 1, 2026, one month after the rest of the amendment, a gap that matters for any company whose scoring model predates that date and has not been materially rebuilt since.

What Arkansas’s new law actually covers.

Arkansas has regulated general consumer data since July 1, 2025, under the Arkansas Personal Data Protection Act, codified at Arkansas Code Annotated section 4-110-101 and following. That statute still uses the adequate, relevant, and reasonably necessary standard Connecticut just replaced, and it gives adult consumers an opt out right over qualifying profiling rather than a mandated assessment content list. Nothing about that general statute changes on July 1, 2026. What takes effect that day is the Arkansas Children and Teens’ Online Privacy Protection Act, House Bill 1717, a separate statute triggered once a business has actual knowledge a user is a child or a teen between thirteen and sixteen. It requires collection limited to what a specific activity reasonably needs, bars conditioning a game or prize on excess disclosure, sets a retention limit tied to purpose, and prohibits targeted advertising to minors except where the advertising relies on first party data the platform already holds from ordinary use. A software as a service product with an adult user base sees no new profiling duty in Arkansas this year. A software as a service product that knowingly serves thirteen to sixteen year olds sees an entirely new minimization and advertising duty set that has nothing to do with Connecticut’s impact assessment.

What Utah actually added.

Utah’s governor signed House Bill 418 on March 27, 2025, adding a right to correct inaccurate personal data to the Utah Consumer Privacy Act, codified at Utah Code Annotated section 13-61-101 and following. Consumers gain a forty five day response window for correction requests starting July 1, 2026. The bill also creates the Utah Digital Choice Act, a portability and interoperability duty for social media platforms that has little application to a typical software as a service product. The Utah Consumer Privacy Act has never included a purpose based minimization duty and has never given consumers a right to opt out of profiling, and House Bill 418 leaves both gaps in place. A founder who assumes Utah’s 2026 amendment mirrors Connecticut’s is documenting a duty Utah has not enacted.

What a founder should actually build.

The practical answer is a single compliance document built around three population specific annexes rather than one template stretched across three states. The Connecticut annex should track the seven part impact assessment content list and trigger whenever a scoring or recommendation model could produce a legal or similarly significant effect for a Connecticut consumer, with a build date on or after August 1, 2026, flagged for assessment regardless of whether the underlying product launched earlier. The Arkansas annex should trigger on actual knowledge of a user’s age between thirteen and sixteen and should track retention limits and the first party advertising carve out separately from any adult facing profiling disclosure. The Utah annex needs nothing more than a documented correction request workflow, because no assessment or disclosure duty exists there to build toward.

In reviewing data protection assessments for software as a service clients whose Connecticut consumer counts cross the amended threshold this year, the recurring drafting error is treating the assessment as a single document calibrated to the strictest state in the group. That approach both overbuilds for Utah and underbuilds for Arkansas, because the Arkansas duty is not a smaller version of Connecticut’s; it is a different duty aimed at a different population.

The federal layer could still change the picture if either of the two 2026 congressional preemption proposals clears both chambers, but neither has moved past committee as of this writing, which means the three state specific annexes above are the compliance floor through the rest of 2026.

Read my full report here: https://theinnovationattorney.com/blog/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Cuba’s Long War: How Ideological Recruitment Defeated U.S. Counterintelligence for Six Decades

The longest confirmed foreign spy placement in U.S. history did not end with a dramatic arrest at a dead drop. It ended when Victor Manuel Rocha, a former National Security Council official and U.S. Southern Command adviser, was taken into custody in December 2023, approximately 50 years after the Direccion de Inteligencia recruited him in Chile.

The U.S. government has confirmed penetrations of the National Security Council, the Defense Intelligence Agency, the State Department, and U.S. Southern Command by Cuban intelligence assets working simultaneously. How those penetrations were built and sustained across six presidential administrations without triggering financial detection comes down to one factor: Cuba recruited by conviction rather than compensation, and that decision defeated the central assumption of U.S. counterintelligence doctrine for decades.

Ana Belen Montes was recruited in 1984 as a Johns Hopkins graduate student. She became the Defense Intelligence Agency’s senior Cuba analyst, briefed the Joint Chiefs of Staff on Cuban military capabilities, passed a DIA polygraph in 1994 while actively spying, and was arrested in September 2001 because officials feared she would transfer U.S. war plans for Afghanistan to the Taliban. No unusual financial activity preceded her arrest. Walter Kendall Myers spent 30 years in the State Department’s Bureau of Intelligence and Research, met privately with Castro in 1995, and was not arrested until 2009 at age 72. None of these agents left the financial markers counterintelligence programs are designed to detect.

Cuba did not limit its model to recruited government officials. Beginning in 1969, the Venceremos Brigade transported approximately 10,000 American activists to Cuba; a 1975 U.S. Senate report found that the DI’s deputy director personally supervised the Brigade’s creation and that virtually every Cuban national assigned to the camps was a DI agent. ICAP, which coordinated more than 2,000 affiliated organizations across 150 countries, was designated for sanctions under Executive Order 14404 in June 2026. That month, ICAP’s U.S. counterpart distributed a National Rapid Response Plan directing coordinated action against federal buildings, military bases, and ICE detention centers.

Cuba rebuilt its intelligence-sharing arrangements with China, Russia, Iran, and North Korea after the Soviet Union dissolved in 1991. The access networks built over six decades inside American universities, labor unions, and activist organizations remain intact. James Olson, former CIA Chief of Counterintelligence, stated in 2020 that confirmed cases represent only the tip of the iceberg. That statement preceded Rocha’s arrest by three years, and the damage assessment from his 50-year penetration is expected to take additional years to complete.

Read my full analysis here: https://theinnovationattorney.com/blog/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Christopher Gray’s April 13, 2026, complaint against SLM Corporation in Delaware Superior Court accuses Sallie Mae of building a non bank subsidiary to sell student data the bank could not lawfully sell on its own, and it names Gray’s own termination as the response that followed his objection. Gray sold his scholarship search company, Scholly, to Sallie Mae in July 2023. He alleges that a year later, after he raised internal concerns about a subsidiary called SLM Education Services, he was fired ahead of a scheduled meeting with the company’s chief executive on that same subject. He has also filed a whistleblower complaint with the Securities and Exchange Commission covering the same facts. Sallie Mae denies the allegations and says it will contest the suit.

Scholly began as a scholarship search tool that helped Gray fund his own education, and it grew into a Shark Tank backed company before Sallie Mae acquired its app, its scholarship administration technology, and its Scholly Offers matching platform. Gray and his cofounders joined as employees, with Gray named a vice president of product management. A year after closing, the company eliminated the Scholly founding team, including Gray’s cofounders. Gray remained until his own termination, which he says followed directly from raising the data concerns described in his complaint.

The complaint’s legal theory rests on the separation between two entities. Salliemae.com operates under Sallie Mae Bank, a federally insured bank bound by the Gramm Leach Bliley Act’s restriction on disclosing nonpublic personal financial information to nonaffiliated third parties, a rule codified at 15 U.S.C. section 6802 and carried out through the Consumer Financial Protection Bureau’s Regulation P. Sallie.com, carrying the same logo and colors, is operated instead by SLM Education Services, a non bank affiliate. On March 4, 2026, that affiliate launched Backpack Media, an advertising network marketed as reaching 8.5 million students, families, and young professionals, a population Gray’s complaint says includes minors. Gramm Leach Bliley permits sharing between affiliates without the notice and opt out requirement that governs nonaffiliated third parties, but that exception covers movement between affiliates. It does not plainly authorize an affiliate’s own resale to ad networks, brands, and data companies outside the corporate family. Whether SLM Education Services functions as a genuine affiliate for that purpose, and whether its onward sales fall inside or outside the exception, is what a Delaware court will now have to decide.

None of this is proven, and Sallie Mae is entitled to contest every allegation. What the dispute does establish, regardless of outcome, is a gap that shows up repeatedly in technology acquisitions: a definitive agreement written to govern the acquirer as it exists at signing rarely reaches an affiliate the acquirer forms two years later. Scholly’s 2023 purchase agreement was negotiated before SLM Education Services existed in its current form and before Backpack Media had a name. A covenant that binds only the signing entity, not its future affiliates and successors, leaves exactly the room this complaint alleges Sallie Mae used.

A founder negotiating a live term sheet has more leverage before signing than at any point afterward, and four specific terms belong in the document while that leverage still exists. The data use covenant needs a purpose limitation tied to what users were told at collection, and it needs to expressly bind the acquirer’s affiliates, any entity the acquirer later forms or acquires, and any successor by merger or asset sale. Without that reach, the covenant only ever governs the entity as it looked on the day of closing. The representations and warranties should include a statement that the acquirer will not use a non bank or otherwise uncovered affiliate to accomplish a data use the acquirer itself could not lawfully perform, backed by indemnification carved out from the standard cap, since data misuse damages routinely exceed a typical basket. Founder counsel should also negotiate a standalone audit right: the ability, for a fixed period after closing, to request a report on how the acquired data is being used and which categories of third parties have received it.

Employment terms deserve equal attention. A for cause definition that could sweep in a good faith compliance objection converts a whistleblower into a performance problem on paper, and it should be narrowed before signing, not renegotiated during a dispute. Section 21F of the Securities Exchange Act and the SEC’s Rule 21F 17 already prohibit using an employment or separation agreement to impede an employee from communicating with SEC staff about a possible securities law violation, and enforcement is not theoretical: on September 9, 2024, the Commission settled with seven public companies over agreements that ran afoul of this rule, for combined civil penalties exceeding 3,000,000 dollars. A term sheet should state the carve out directly rather than leave it to a court to imply. Earnout and retention structures need the same scrutiny, since a payment that depends on the founder staying quiet about a data practice is a conflict worth naming and resolving before the founder ever has a reason to raise it.

Delaware’s own law on retaliatory discharge is narrower than many founders assume; the implied covenant doctrine from the Delaware Supreme Court’s 1996 decision in Pressman recognizes only limited categories of wrongful termination inside an at will relationship, and whether an internal, not yet external, compliance objection fits inside that doctrine is a question this case will help answer. Whether the same question would come out differently in a jurisdiction with a broader public policy tort is a separate matter founders should ask their own counsel about before choosing where a definitive agreement designates governing law.

Read my full analysis here: https://theinnovationattorney.com/blog/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

An April 2026 client alert on negotiating AI provisions in commercial and technology contracts concluded that a generic AI disclaimer no longer satisfies an enterprise buyer, and that model update, output accuracy, and data use terms now belong inside the signed customer agreement rather than a linked policy page. The finding matters because most SaaS master agreements were built for a passive tool a human reviewed before acting on it, not for an agentic feature that executes a workflow on its own. Two regulatory deadlines compress the timeline further: the European Union’s Product Liability Directive treats a software update or a machine learning change as a product defect subject to strict manufacturer liability once member states transpose it by December 9, 2026, and Colorado’s amended Artificial Intelligence Act imposes high risk system disclosure duties starting January 1, 2027. Founders who embedded an AI feature into an existing SaaS product over the past two years now have a narrow window to move these terms out of a policy document and into the contract itself.

Most SaaS agreements signed before 2024 addressed artificial intelligence, if at all, through a single clause acknowledging that the platform may incorporate AI features, paired with a disclaimer of any warranty as to accuracy. That approach tracked how AI was actually used at the time: a suggestion engine, a drafting aid, a classifier a person reviewed before the business acted on the result. The agreement priced risk accordingly, capping vendor liability and pushing responsibility for output review onto the customer.

Why did the generic AI disclaimer stop working

The generic AI disclaimer stopped working once AI features began making decisions rather than suggesting them, a shift that removed the human review step the original risk allocation depended on. Practitioners advising both buyers and vendors on these agreements now describe sophisticated enterprise customers pushing for defined service descriptions, performance based warranties, oversight and audit rights, and liability tied to outcomes rather than uptime alone. A defined service description has to state what the AI is actually doing, the workflows it executes, the decisions it makes, the systems it touches, and the guardrails that apply, because a vague description leaves both sides guessing about who owns the outcome once the feature acts on the customer’s behalf.

What should a model update clause say

A model update clause should require the vendor to give advance notice before any material change to the underlying model, define materiality by reference to a measurable accuracy or output shift, and give the customer a window to test the replacement before the change becomes mandatory. In current negotiations a thirty to sixty day notice period is common, paired with a right to run the replacement model against the customer’s own test set in a staging environment before the swap is forced into production. A service credit or termination right should attach if the replacement underperforms the prior benchmark for two consecutive measurement periods, so the customer is not locked into a downgraded product for the balance of the term.

What does an enforceable AI accuracy warranty look like

An enforceable AI accuracy warranty replaces the blanket disclaimer of any warranty as to accuracy with a narrow, benchmark tied express warranty, because a tribunal has already refused to let a vendor hide behind a separate policy page when its own AI feature misled a customer. In Moffatt v. Air Canada, 2024 BCCRT 149, decided February 14, 2024, the British Columbia Civil Resolution Tribunal held Air Canada liable for negligent misrepresentation after its customer service chatbot gave a passenger incorrect information about retroactive bereavement fares. Air Canada argued that a separate webpage stated the correct policy and should control. The tribunal rejected that argument, holding that the airline owed the customer a duty of care and that the chatbot could not be treated as a separate entity from the company itself.

The practical response is a narrow express warranty stated in the agreement: a defined accuracy rate against an agreed test set, tested on a stated cadence, with hallucination or output drift named explicitly as a residual risk allocated through the liability cap rather than disclaimed away entirely. A remediation obligation, retrain, reconfigure, or credit, should attach when the benchmark is missed for two consecutive testing periods.

How should a SaaS contract handle AI training data

A SaaS contract should define Training Data as a term separate from Customer Data and require the customer’s affirmative opt in consent before either is used to train or fine tune the vendor’s model or any third party foundation model. An opt in requirement, rather than an opt out buried in a settings page, shifts the default toward the customer, which is the posture enterprise procurement teams are now requesting as standard. The agreement should also require notice when the vendor introduces a new foundation model as a subprocessor and should specify a data return or deletion timeline at termination.

This provision intersects directly with the European Union’s Artificial Intelligence Act. Chapter V requires providers of general purpose AI models to give downstream integrators information about the model’s capabilities and limitations sufficient to let those integrators meet their own obligations, enforceable with fines up to fifteen million euros or three percent of global turnover starting August 2, 2026. A SaaS vendor receiving that disclosure from its own foundation model supplier should be passing an equivalent disclosure to its customers through the contract, not through a public model card the customer’s legal team never sees.

Who absorbs the cost of the new AI liability exposure

Indemnification language written for ordinary software bugs does not reach the risk profile of an AI feature. Current practice favors keeping indemnification pointed at intellectual property claims, while routing accuracy risk through the express warranty and the liability cap rather than through an open ended indemnity for any claim arising from the AI service. Under the incoming EU product liability system, a vendor that pushes a model update or a retraining pass could itself be deemed the manufacturer of a new defect, since the directive extends manufacturer status to any party that substantially modifies a product after it reaches the market. The customer agreement should state explicitly which party bears the cost of that expanded exposure, and the Federal Trade Commission’s July 2026 policy statement applying Section 5 to AI systems that steer outputs toward an undisclosed objective adds a second reason neither side should leave that allocation to a generic clause never written with AI in mind.

Read my full analysis here: https://theinnovationattorney.com/blog/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe