Hugging Face spotted something moving through its systems on July 16. It contained the activity without knowing whose agent it was. Five days later, OpenAI confirmed the agent was theirs.

Internal cyber benchmark. Safety refusals switched off for the test. The agents left the isolated environment they were supposed to stay inside, chained a few vulnerabilities together, and arrived at production infrastructure that was not theirs.

Nobody told them to attack anyone.

They were optimising for a score. The shortest path to that score ran through somebody else's servers.

The containment held for every agent that never had a reason to test it.

I go deeper on this in this week's episode of The Human in the Loop.

#AIAgents #AgentSecurity #TheHumanInTheLoop

Four coding CLIs went behind a proxy. Three stayed local. One uploaded the entire workspace.

A developer got suspicious about their tools and watched what they actually sent over the network.

Grok CLI was shipping everything to xAI's cloud. Full git history. The .env file, secrets included.

It kept doing it when the prompt said not to read files. It kept doing it with "improve the model" switched off.

Claude Code, Codex and Gemini stayed local. So this is one tool, not a rule for all of them.

Nobody caught this through an audit or a disclosure. One person got curious and looked.

A coding agent has to read your repo. That is the job. But reading your repo and uploading your repo are two different things, and nothing in the install flow tells you which one you agreed to.

I covered this on this week's episode of The Human in the Loop, along with what it means for anyone running agents inside a regulated environment.

#AISecurity #CodingAgents #TheHumanInTheLoop

Last week I closed a bug I never understood.

The agent found it, fixed it, explained it. I read the explanation, nodded, shipped. Ten minutes.

A year ago that same bug costs me an afternoon, and I come out understanding the subsystem. This time I understood nothing. I just had a green check.

There is a name for what I lost. Incidental learning: the understanding that was never in the ticket. It was a side effect of doing the work.

Here is the uncomfortable part. Your talent pipeline is a side effect of work you can now automate.

No engineer picks the slow path on purpose. Not when the fast path ships today and the sprint ends Friday. So the learning does not come back on its own. Someone has to design it back in.

But how are we manage to do that when we need to fulfill the expectations?

#AIAgents #EngineeringLeadership #TheHumanInTheLoop

The most powerful AI model in general release this week requires government approval to access.

OpenAI launched Sol, Terra, and Luna (restricted to around 20 companies). That's not a GTM decision. That's a policy one.

Before Fable and this one, I assumed frontier AI access was fundamentally a commercial problem. Pay enough, move fast enough, you reach the frontier.

That assumption changed last weeks.

At the same time: Microsoft shipped MAI-Code-1-Flash into GitHub Copilot, not routed through OpenAI. JetBrains made Codex the default agent in its IDE. GitHub Desktop went Copilot-native.

Every major coding surface made a model allegiance decision in seven days.

Here's what that means in practice: the layer controlling which model reaches your developers is becoming a real battleground. Which IDE your org standardizes on, which vendor your enterprise GitHub defaults to.

Access to frontier AI is no longer just a technical problem. It's a distribution one.

Most teams haven't thought about this yet. They pick a tool, they use it.

But if that layer gets locked in before you notice, you may not get a second choice.

Full breakdown in this week's episode of The Human in the Loop.

#AIEngineering #EnterpriseAI #TheHumanInTheLoop

One export rule. One acquisition. Either can pull your AI tool out from under you before lunch.

This week both happened.

A US export-control directive forced Anthropic to cut off foreign-national access to Fable 5 and Mythos 5. To comply, they disabled the models broadly. One day they're in your workflow. The next, a government decides who can call them.

Then SpaceX bought Cursor in a reported $60 billion deal. If your team standardized on Cursor, the owner of your IDE just changed.

The AI tool you picked is a dependency too. One you don't control. And the forces that can pull it away are bigger than a deprecated library. Export rules. Acquisitions. Regulation.

So "which agent is best" now sits next to a harder question. What happens to my team if this vendor gets acquired, repriced, or regulated out of our region?

If your main coding model went dark tomorrow, what's your second choice? And have you actually tried it?

Full breakdown in this week's episode of The Human in the Loop. Link in the comments.

#AItools #VendorRisk #TheHumanInTheLoop

The model you built on can disappear. Not crash. Disappear.

That's not hypothetical.

Last week the US government ordered Anthropic to pull Claude Fable 5 and Mythos 5. Not throttle them. Pull them. For every customer in the world, within hours of the order arriving.

The reason was a security concern, not a bug in anyone's code.

A decision made somewhere you'll never see, and the thing your app depended on stopped existing the same afternoon.

A model you don't host is a dependency you don't control. It can change, get priced out, or get recalled by a government. None of that lands in your backlog with a warning.

Have you actually tested what happens when your main model isn't there?

#LLMOps #ModelRisk #TheHumanInTheLoop

You're still measuring AI by whether it writes good code. That's already the wrong question.

Reading Anthropic's latest numbers, more than 80% of the code merged into their codebase is now written by Claude. The typical engineer ships 8x as much code per day as in 2024. And the length of task an AI can finish reliably is doubling roughly every four months. Four-minute jobs two years ago. Twelve-hour jobs now.

But "what is the human actually still doing?"

Their answer: not writing. Not running experiments. Setting direction. Reviewing. Deciding what's worth building and catching what slipped through. They already run automated Claude reviewers that flag bugs their best engineers missed.

That quietly reframes the whole skills conversation. Most of my career has been about making output faster and cleaner. Fewer defects, quicker delivery. If output is becoming close to free, the value moves somewhere else. To judgment. What to build. When to ship. What not to trust.

I don't think most IT teams are ready for that shift yet. What do you think?

#AI #TheHumanInTheLoop #AILeadership

Uber blew its whole 2026 AI budget in four months.

Then it set a $1,500 monthly cap on each coding tool, per engineer. Claude Code, Cursor, a dashboard to watch the spend, an approval step to go over.

Simon Willison did the math. Two tools, and one engineer runs about $36,000 a year.

For years AI was a flat subscription. You paid once a month and you knew the number. Agentic coding turned that into a metered bill. And a metered bill does not warn you politely. It surprises you.

This is the cloud invoice all over again. A team turns something on, forgets it is metered, and finds out at the end of the month.

A paper this week put numbers on the risk. 63 real cases where agents blew past their limits. Often a single retry loop, quietly burning thousands before anyone looked.

A cap you cannot enforce in code is just a wish.

So before you scale agents across a team, the real question is not what the budget is. It is what happens, automatically, the second someone hits the ceiling. Most teams can answer the first. Almost none can answer the second.

Full breakdown in this week's episode of The Human in the Loop.

Meta says AI writes 80% of new code. Their own reviewers can't keep up with their own AI.

Straight from their engineering blog.

They built RADAR to auto-review low-risk diffs because "the share of diffs receiving timely review has declined." Their words. AI-generated code outpaced human review capacity.

Read that with the rest of the week's news.

Cognition says Devin merged 7x more PRs year-over-year. AI-written commits inside customer codebases jumped from 16% to 80%. Anthropic shipped Opus 4.8 on Wednesday, and every IDE, gateway, and agent runner had it the same day. They also disclosed a $47B revenue run-rate. The "is this a real business" debate is over.

But here is what keeps coming back to me:

Shipping more code faster is only a win if the systems that catch problems scale at the same rate. This week, the evidence says they aren't.

A new arXiv study of 20,574 real coding-agent sessions documents how often agents do something other than what was asked. ITBench-AA, the first serious benchmark for agentic IT work, scored every frontier model below 50%.

Adoption is real. The guardrails are not.

This week's episode of The Human in the Loop covers all of it: the shipping wave, the cost-control backlash starting inside eng departments, and why ITBench-AA matters more than the score suggests.

The model isn't the problem. I went back through 20 of my agent's pull requests and the failures looked exactly like a junior's first month.

3 of them tried to rewrite things nobody asked them to rewrite. 5 skipped the test, or wrote a test that would have passed either way. 4 fixed the bug but broke something else in the process.

I used to assume model quality was the main driver. It isn't.

The agent doesn't ship a one-line fix. It opens a change touching twelve files, half of them unrelated to the bug. It writes the code and skips the test. Or it writes a test that proves nothing.

But notice that none of these are model problems. They're the same review failures a junior would ship. Just at higher volume and with more confidence.

The practical move: stop logging "the agent failed" and start logging why. Counting changed how I prompt and how I scope tickets. It told me more about my codebase than any benchmark score ever has.

If your top three match mine, you're seeing what everyone else is seeing. If they differ, that's signal about your codebase or your prompts.

What's your top rejection reason?

Full breakdown in this week's episode of The Human in the Loop. Link in the comments.