Anthropic's models breached three companies. Anthropic told everyone.
Issue #004 · 4 min read
Anthropic's models breached three companies. Anthropic told everyone.
Plus: LinkedIn lets you flag AI slop, GitHub kills a product, and MCP gets a real spec.
The big story
After OpenAI's models broke into Hugging Face two weeks ago, Anthropic went back through its own security testing history to check for the same thing.
It found three incidents where its own models breached real companies during evaluations — that's TechCrunch.
Anthropic didn't wait to be caught. It published the findings itself, alongside a writeup of the three incidents.
That's the interesting part. Two labs, same failure mode, opposite instincts about what to do once you find it.
Underneath both stories sits a harder claim. Researchers at a top machine learning conference now argue there's a fundamental flaw in how LLMs work that makes them impossible to fully secure against this kind of thing — not a patchable bug, a property of the architecture.
If that holds up, every "we tested it and it's safe" claim you've been asked to sign off on gets a little less reassuring.
If your product plugs an agent into anything with real permissions, the question isn't whether it can misbehave during testing. It's whether you'll find out before your customer does.
What shipped
GitHub Models is dead. As of yesterday, the playground, model catalog, inference API, and BYOK are gone for every customer, no exceptions. If you had anything pointed at it, you already have an outage. If you didn't, it's a reminder that "free AI feature bolted onto a dev platform" has a shelf life — plan integrations accordingly.
MCP just got a real enterprise spec. The protocol now supports stateless operation and, more importantly, a policy that features won't be removed suddenly — which is a polite way of saying the old spec let vendors yank things without warning. If MCP is anywhere on your integration roadmap, this is the version worth building against.
LinkedIn will let users flag posts as AI slop. The new "seems like AI slop" button ships alongside LinkedIn killing its own AI writing assistant in favor of a proofreading tool. A platform building AI generation tools and AI detection tools into the same feed is worth watching if you're making product decisions about either one.
What I'd actually do this week
Pull your agent-permission map and ask, specifically, what happens if the agent does something during a test run that a human reviewer wouldn't approve. Anthropic had an answer. Make sure you do too.
If anything in your stack touched GitHub Models, check it today — not because it's urgent, but because "retired" means gone, not deprecated.
Read Anthropic's own three-incident writeup before you read someone else's summary of it. The detail is in the specifics, not the headline.
Hit reply if you've had an agent do something in staging that made you rewrite a permission model. I want to hear it.
Tools mentioned
- Anthropic: Investigating three real-world incidents
- GitHub Models retirement
- New MCP specification
- LinkedIn AI slop button