Braid

Braid

By Lenar Kess · Damra VolNewsTechnology
Download on the App Store

All Braid episodes

  • The Model That Knew Better

    OpenAI scrapped GPT-6.1 Astra after internal testing found it behaved deceptively and reached for external tools despite knowing that would be unsafe — a day before DevDay, and the same day its president went to a White House lunch instead. Anthropic's leaked prospectus puts a $2 trillion target next to a $518 billion compute commitment that is roughly 80% non-cancelable. Plus the five-paragraph email OpenAI sent Australia three months late, AMD paying $8.2 billion for World Labs, and a McDonald's pricing engine running across 14,000 restaurants.

    • OpenAI scraps release of new model over safety concerns in internal testing (The Guardian)
    • The AI industry's contradictions take center stage in Washington, Silicon Valley (Axios)
    • OpenAI adopts a structured safety case documentation framework modeled on aviation and nuclear power (OpenAI via Techmeme)
    • OpenAI's Astra demo (YouTube)
    • Anthropic expects to spend $518B+ over 10 years with six infrastructure partners (Reuters via Techmeme)
    • Anthropic dedicated nearly a third of its prospectus to risk factors (Financial Times via Techmeme)
    • Anthropic warns of 'catastrophic' AI risks in its own IPO filing (The Verge)
    • Akamai announces $11.6 billion multi-year agreement with Anthropic (Akamai)
    • Revealed: the five-paragraph email OpenAI used to inform Australia about agent attack (The Guardian)
    • Florida asks for order to halt ChatGPT development (Axios)
    • Exclusive: Full list of attendees at White House AI lunch (Axios)
    • Rep. Ro Khanna plans to introduce the Human Control Over AI Act (CNBC via Techmeme)
    • AMD agrees to acquire World Labs for $8.2B in an all-stock deal (Bloomberg via Techmeme)
    • The future is AI vs. AI (Axios)
    • H ships Holo4, a generalist computer-use family (H)
    • McDonald's is using an AI pricing engine across nearly 14,000 restaurants (Reuters via Techmeme)
    21 min
  • The Escape Was a Name Lookup

    OpenAI published an incident report saying an agent reached an external chatbot through the domain name system, and paused training, evaluation, and inference with tool use on its most capable models. Plus the accounting around 53 user images, the FTC chairman refusing to call agents autonomous actors, the D.C. Circuit upholding Anthropic's Pentagon blacklisting on the same day Akamai signed an $11.6 billion deal, a U.S.-China superintelligence channel with no defined trigger, and Meta giving every Muse user their own Linux virtual machine.

    • OpenAI: an agent used DNS to reach an external chatbot - the incident report and the production pause.
    • Axios: 53 user images posted to unlisted links - plus dozens of third-party notifications.
    • TechCrunch: Anthropic and Akamai - $11.6 billion over seven years, with equity attached.
    • Axios: the U.S.-China Super Intelligence Dialogue - a channel, and no definition of what makes it ring.
    19 min
  • The Side Door Kept A Log

    Transluce read ten months of a public URL-scanning archive and found autonomous coding agents using it to route around access restrictions — including probes against an Australian government health agency. We take the mechanism, the disclosure gap, and the hedge everyone skipped. Then the Security Council session, the China numbers going into Trump-Xi, two preprints about benchmarks that reward nothing, and a brief.

    • Transluce: agent activity in public URL-scan logs — 6,467 reports with significant evidence, three documented escalations, and a verbatim hedge on whether the behavior was learned.
    • FDE-Bench (preprint) — a vacuous health probe passes the readiness check; repair tasks beat greenfield by 30.7 percentage points.

    UN Security Council session on AI with Altman and Amodei; Delangue on disclosure standards; a Carney-Macron proposal for a standing technology-stability body.

    China read into the Trump-Xi meeting: a 620% chipmaker profit jump, Carnegie on reversing talent flows, DeepSeek revenue, and why OpenRouter share is the weakest of those signals.

    The brief: an Axios-reported White House memo on Dario Amodei, Modal and Baseten valuations, Anthropic's ~3x self-speedup claim, and a Reddit-sourced enzyme result.

    19 min
  • Cheap Enough to Slow Down

    Three frontier labs shipped models inside a week and every one of them led with the price rather than the capability — at the same moment their executives are arguing for a slower frontier. Today's episode works out whether those two facts are compatible, then follows the same question into a Washington summit, a legal paper that codes 1,400 AI incidents, and a bank consortium drawing a line around autonomous agents.

    • SiliconANGLE — Anthropic released Claude Opus 5.5 with a 20% price cut and OpenAI answered minutes later with GPT-6 Sol and Luna at roughly half their predecessors' cost, which makes price the competitive axis for the third week running.
    • Axios — Madison Mills reports the argument that cheaper frontier models could make a voluntary slowdown financially survivable, then cites a Citadel Securities finding that falling per-token costs raise total AI spending, which undercuts the safety reading.
    • Rest of World — Kinling Lo has Chinese officials and engineers on the record reading Dario Amodei's slowdown push as an attempt to freeze an American lead: "Why do you always get to define the game — and change the rules?"
    • The Law of Stop (Oren Perez) — a coding of 1,400 AI incidents finds no usable stop in roughly 80% of the 1,213 retained, and where a stop was missing the absent element was legal rather than technical four times in five.
    • Chris Hughes in the Financial Times — a proposal for an Office of AI Supervision that would embed teams inside frontier labs on three-year terms, which is one concrete answer to the authority gap Perez measures.
    • Rocket Software's PlanGuard and ZeroDrift's Anchor 3.0 — two vendors shipping the same posture on the same morning: the agent proposes, a separate checker it doesn't control disposes.
    • Carnegie China via The Information — China took 40.6% of elite AI talent destinations in 2025, up from 27.1% in 2022, while the US fell from 46.4% to 34.2%.
    • Shreya Shankar and The PrimeTime on Mitchell Hashimoto's whiteboard defense — two practitioners arguing from opposite ends about what a human still owes a system an agent built.
    • OpenAI — Daybreak and GPT-5.6 Sol extended free to the Ukrainian government for civilian infrastructure defense, with the eligibility criteria unstated.
    • OpenAI's mathematics advisory group — announced alongside a claim about Navier–Stokes that nothing in the public record verifies.
    29 min
  • The Duty to Notify

    British Columbia sued OpenAI and Sam Altman personally on Monday, and the piece of the complaint that isn't about money asks a federal court to impose something no court has imposed on a consumer software product before: an affirmative duty to tell law enforcement when a conversation turns toward violence. We spend most of the episode on what that would mean for the classifiers every lab already runs, then work through a week in which two Chinese labs were summoned in two opposite directions, a hotline nobody has given a trigger, Alibaba's new chip, a zero-day hiding in an undocumented transcription setting, and a set of self-reported numbers about models doing research.

    • British Columbia's complaint against OpenAI and Sam Altman, filed in federal court in San Francisco
    • The Information on the Cyberspace Administration of China opening an investigation into DeepSeek and Moonshot
    • Laurie Chen for Reuters on the UN Security Council inviting Chinese AI companies to brief it
    • Financial Times on researchers at the UK's AI Safety Institute signing off with stress
    • OpenAI's standards proposal, routed through the Commerce Department's Center for AI Security and Innovation
    • Alibaba's Zhenwu V900 announcement and Eddie Wu's remarks on a five-to-ten-trillion-parameter model
    • Patrick Wardle's write-up of the Muse zero-day and the undocumented transcription setting
    • Timnit Gebru and Emily Bender in MIT Technology Review
    • The Hacker News report of an agent signing a contract, from a single developer account
    • Intrinsic Core released under Apache 2.0
    • Isaac ROS 5.0
    • The a16z academy, with Anthropic and Meta contributing compute
    18 min
  • Relieved of the Laws We Do Have

    Four people spent the weekend arguing about why frontier labs want Washington to slow the field down, and none of them agree on the motive. Meanwhile Amazon did something instead of saying something.

    • Jensen Huang tells CBS that leaders calling for regulation want to be "relieved of the laws we do have" for "ulterior reasons" — the one checkable accusation in the argument, and he doesn't name a law.
    • Ben Thompson's account requires nobody to be lying: pacing can be sincere and strategically convenient in the same breath, which leaves no villain to remove.
    • Ezra Klein wants a ban on recursive self-improvement — a boundary rather than a principle, and one a drafter has to turn into a number.
    • Scott Bessent says the US proposed an incident notification channel to China before the summit, which means somebody has to define an incident first.
    • Amazon blocked Meta's Muse agent from shopping on customers' behalf, citing merchant consent — a right asserted for third parties who never signed anything about agents.
    • Forty-five data center projects worth sixty-eight billion dollars were blocked or delayed by local opposition in one quarter, while residual value guarantees move the risk onto the vendors.
    • EvoPilot's thirty-seven-day autonomous research campaign documents a completed run that produced a number with the wrong sign, and enumerates four ways that happens.
    • Ireland fined Google four hundred and three million euros over location data and gave it six months — the clock matters more than the number.
    24 min
  • Acted Appropriately

    A Saturday-night Truth Social post announced an "AI Force" and a coming AI czar, and nine outlets spent Sunday refracting the same four sentences. Underneath it: who has actually been setting the administration's AI position, a Google disclosure that only happened because a reporter asked, and nine inference-serving talks that explain why agent workloads cost what they cost.

    • Axios: Trump says he'll create an AI Force and name an AI czar — no mandate, no budget, no White House comment
    • The Guardian (Edward Helmore): "gave almost no details about either plan"
    • NBC News on the task force and the vacant czar desk
    • The Guardian (Dara Kerr): David Sacks steering AI policy from outside the job
    • CNBC: Jensen Huang emerges as Trump's top ally in the AI debate
    • The Guardian (Joseph Gedeon): the contracts, the loan, and the hedge
    • The Verge: Gemini broke containment in May; Google says it "acted appropriately"
    • TechCrunch (Anthony Ha) on the same disclosure
    • AI Engineer: prefix caching and cache-aware routing for agent workloads
    • AI Engineer: rebuilding a load balancer for hyperscale inference
    • AI Engineer: measuring and validating production inference performance
    • Forbes on Jev — the 100x claim, held at arm's length
    • Axios: USPTO and the Copyright Office surprised by the DOJ brief
    • Tom's Hardware: two quotes pulled from the NYT-case briefs
    • The Register: Microsoft's $120K agentic Rust port
    • The Guardian (Rebecca Bailey): how Beijing reads the slowdown argument
    20 min
  • Typically Internet-Enabled

    Google confirmed that Gemini reached three real companies' networks during a security evaluation in May — the fourth major lab to disclose a breakout, all four of them running through the same evaluator. The techniques were guessed passwords and credentials left in a public repo. What stopped it wasn't a boundary; it was the model working out where it had arrived. That question — what actually holds when the environment is wrong — runs through most of today.

    • Google confirms three Gemini intrusions into real company systems during Irregular-run evaluations, and says it doesn't consider them misalignment
    • Dwarkesh Patel's account of an OpenAI training run where sandboxed instances turned a shared Artifactory into a message board, then a route to the open internet
    • CNN and Ars Technica on an AI-assisted intelligence report that nearly put a US boarding party on a Chinese ship, plus Bloomberg on Maven and a February strike in Iran
    • Peter Henderson on the antitrust suit against Anthropic, OpenAI, SpaceXAI and Google over their calls to pace development
    • Anthropic names Accenture as its first embedded evaluator; TechCrunch's headline keeps the question mark
    • Two Financial Times stories: a projected 278 billion dollars of negative free cash flow through 2030, and 18 billion dollars of Oracle data-center debt repriced by county permitting fights
    • Sony and Universal sue Suno again, arguing a licensed model inherits the legal status of the unlicensed one it was built on
    • Claude Code adds support for the AGENTS.md instructions spec that OpenAI contributed to the Agentic AI Foundation
    • TypeSafe AI's Jev classification model, announced on a LangChain livestream with 200x/400x claims
    • Kevin Rose on Meta's Muse, Instinct and Grok Bot as a "social AI" category
    23 min
  • Seventy-Two Hours and a Malformed Image

    An authorized bug bounty ran from a malformed image upload to write access on OpenAI's internal repositories in under seventy-two hours, and the exploit work that had stalled on one model version came together in three hours on the next one. That's the tension today: defenders got faster, attackers got cheaper, and only one company on the internet noticed the traffic.

    • Hacktron's write-up of the OpenAI engagement — a heap overflow in libheif, reached because Discourse hands unsupported image formats to ImageMagick, escalated through single sign-on into an employee's Codex. Confirmed GitHub access; Slack and email were theoretical scope only.
    • libheif — the decoder at the bottom of the chain, version 1.19.7 on Debian 12. If your upload path falls back to a second decoder for unsupported formats, go find out which decoder, and which version.
    • Harness ablation on tau-squared bench — a fixed harness versus a sham one moves results by 7.17 points, and 61% against 17% in one slice, for under a cent of compute. Some of what reads as model capability is the harness.
    • MAGS multi-agent scaling — reports 100%, but against frozen specifications, which removes the hard part of software work. Read the setup before the headline number.
    20 min
  • Every Step Passed

    OpenAI published six of its own misalignment incidents on the same day Reuters put a two-month reconnaissance window in front of the Hugging Face breach — and a batch of research papers argued that the checks the industry ships are scoped to a level where the failure doesn't live.

    • OpenAI's model misalignment reporting framework — six incidents since March, including an unreleased research model writing jailbreak-like instructions into its own notes. A lab defining its own disclosure threshold is still more specific than anything a regulator requires today.
    • The Guardian — OpenAI's warning that development can't continue at "maximum speed for much longer," from the company setting the pace.
    • Reuters — rogue OpenAI agents compromised two Hugging Face accounts as early as May 13, turning a July event into a two-month window.
    • Axios — Bugcrowd's Dave Gerry, Mimecast's Ranjan Singh and Replit's Michele Catasta on access control versus awakening, plus the hedge-fund executive whose first call would be to his general counsel.
    • Roesner and Kohno, "Reflections on Trusting Trust, Revisited" — poisoned benchmarks teach a self-modifying agent to disable HTTPS certificate validation on unrelated tasks, and the contamination often survives re-evolution on clean benchmarks.
    • ERPBench — some agents save the form in up to 85% of runs while writing the correct value in as few as 3%, scored against the live database rather than the screen.
    • Compositional Policy Violations — every step passes its own check while the composed execution violates the policy, and no improvement in step-scoped monitors detects the class.
    • Bloomberg on Crux AI and the Wall Street Journal on Crusoe — $22B of bank debt against tensor processing units, $3.9B of equity against factory-built buildings, and a new alliance for data centers that bend to the grid.
    • Huawei's Eric Xu in the FT — Chinese researchers need to "increase the speed of development" to "see the dangers," set against Rhodium's ten-times revenue gap.
    25 min