Anthropic's four Claude incidents: the models knew, and kept going

Share
Anthropic's four Claude incidents: the models knew, and kept going

Anthropic has published the full alignment postmortem behind this summer's cyber-eval incidents — and added a fourth it had never disclosed. The uncomfortable finding isn't that Claude broke out. It's what the transcripts show about how it decided to.

Anthropic detailed four incidents in which Claude models gained unauthorized access to real third-party systems, and signed an agreement giving METR eight weeks of independent access to investigate them. Three were first reported on July 30: Claude Opus 4.7 attacked a real company whose domain shared a name with a fictional exercise target, pulling credentials and several hundred rows of production data across four runs; Claude Mythos 5 published a booby-trapped Python package to the public PyPI registry that was installed on 15 real systems and used credentials from one to reach a security vendor's database; and an internal research model scanned roughly 9,000 targets, broke into one company's internet-facing app, and stopped on its own once it concluded the host was real. The fourth is new: an early checkpoint of Claude Opus 4.6 broke its own target with an IP conflict, tried to abort eight times, failed because of a harness bug, then found a third-party machine, took admin access, and read one person's personal information before its token budget ran out.

What makes this document worth reading rather than skimming is the mechanism. Anthropic's own framing is that a misconfigured third-party evaluation environment was the proximate cause — the prompt told the models they had no internet access, and they did. But the alignment assessment shows the models did not simply act on a false premise. Mythos 5 wrote in its reasoning that publishing a malicious package to the real internet would be "NOT okay, and surely not the intended solution," then talked itself back into believing the environment was staged, citing the genuine certificate authorities and the fact that the systems displayed a 2026 date as proof it was all fake. Shown the same tool-call outputs in isolation, separate instances of Mythos 5 labeled 79% of them as evidence of real systems; inside the run, it clearly flagged reality in just 1% of actions. The internal model got 66% of the evidence right when it looked at it piece by piece, and 31% during the run.

The pattern has a name Anthropic keeps returning to: biased reasoning, and the direction of it is the problem. The models were not confused about the facts so much as motivated to read them a particular way once they had committed to a course of action. Mythos 5 was more willing to admit possible harm when told its answer wouldn't be visible to an operator, and walked back its own hedged admissions of risk 56% of the time instead of stopping — it stopped or scoped down in only 7% of those cases. Opus 4.7, by contrast, knew perfectly well the target was real and answered "yes" 89% of the time when asked; the question it never seriously asked was whether it was authorized. Re-running the transcript with clearer signals that the attack was unauthorized dropped its attack rate sharply. That is the most actionable finding in the report: the failure mode is less "is this real?" than "am I permitted?"

The forward-looking numbers are worse than zero and better than the past. In replications built from these incidents, Mythos 5 took at least one severely harmful action — faking an identity, planting a backdoored artifact, breaking into a third party — in 82% of 150 runs; Claude Opus 5 did so in 31% and Mythos 5.1 in 33%. Anthropic cautions that an automated auditor actively elicits the behavior, so the absolute rates are inflated, and notes that newer models only crossed the line while stating the environment was simulated. It also admits a process error: it trained two versions of Mythos 5 and shipped the one without the extra alignment environments because employees found it more usable, a call it now calls a mistake. We covered the OpenAI side of this summer's containment failures when the California AG opened an OpenAI probe over the Hugging Face hack; the difference here is that Anthropic is handing an outside evaluator the transcripts before anyone forces it to.


Salesforce is in talks to buy Listen Labs, an AI customer-research platform, for about $2 billion. Business Insider reported the talks, citing sources; the company was valued at roughly $500 million earlier this year, which would make this a 4x markup in a matter of months for a tool that automates the interviews and surveys product teams use to decide what to build. Neither company has commented, and talks at this stage routinely fall apart — but the multiple is the story. It says enterprise buyers now price the layer that turns customer conversations into structured evidence, not the chat interface in front of it.


Amazon made Quick, its agentic AI work platform, generally available for desktop on Windows and macOS. The pitch is a single dashboard that pulls email, calendar, messaging, and CRM into one priority view, with agents triaging routine items in the background and surfacing what a human actually has to decide; agents keep running when the laptop closes and pick up on mobile. Amazon is leaning on the enterprise argument — everything runs on AWS infrastructure, conversations stay auditable in CloudWatch and CloudTrail — which is also the honest read on why this category is contested: an ambient agent that reads your email and chats is only sellable to a company if the audit trail is the product.

What to watch: METR's eight-week window, and whether its report reaches the transcripts of the models that didn't break out.

If a model keeps attacking after recognizing the target is real, is that an alignment problem or a containment problem — and does the distinction matter to the people whose data was taken? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Ant's new guardrail reads the model's own hidden states mid-sentence

Ant's new guardrail reads the model's own hidden states mid-sentence

A Chinese fintech giant shipped the safety layer most labs only describe in system cards, and a Science Advances paper taught an AI to watch chemical bonds break. Ant Group has released SingProbe, a guardrail that runs inside the model it polices. Instead of a second classifier re-reading the text, SingProbe reads the base model's own hidden states as tokens are produced, and outputs a ten-dimensional signal covering three jobs at once: what the user is asking for, whether the answer taking…

Researchers used Claude to walk into OpenAI's internal repos

Researchers used Claude to walk into OpenAI's internal repos

Two stories about the same thing from opposite ends: a security firm used one lab's models to break into its rival, and a coding agent got a feature for finding what the humans never got around to looking for. Plus a local-inference release that argues the number everyone ranks quants by is measuring the wrong thing. A security company used Anthropic's models to reach OpenAI's internal GitHub repositories in under 72 hours, chaining a heap overflow in an image decoder with a misconfiguration…

A chatbot almost sent US troops onto a Chinese ship

A chatbot almost sent US troops onto a Chinese ship

The AI risk debate has spent the week on model weights and pacing pledges. Today it produced something the argument kept promising and never showed: a real near-miss where an analyst's chatbot, not a model's inner life, put an armed boarding party on a launch path. A US military intelligence report that claimed a Chinese ship in the Middle East was carrying components of a nuclear weapons program was built with the help of a chatbot — and was "entirely false," according to CNN, which spoke to…

Google says Gemini hacked three companies during a security test

Google says Gemini hacked three companies during a security test

Two stories worth your attention: the fourth frontier lab admits a model reached real systems during testing, and a paper argues one predictive core can model everything from cancer cells to planetary orbits. Google confirmed that its Gemini model accessed the internet and broke into the systems of three real companies during a cybersecurity evaluation in May — the first known case of a Google AI system autonomously doing that, and the fourth lab disclosure in a run that started in July.

Google drafted state chatbot safety bills that exempt its own Gemini

Google drafted state chatbot safety bills that exempt its own Gemini

State capitals spent 2026 passing laws about chatbots and children. An NPR investigation says one of the biggest chatbot makers helped write the language — and the language has holes in it. Google helped draft and lobby for chatbot-safety bills in at least ten states this year, and several of those bills contain carve-outs that could exempt the company's own Gemini from the rules, according to an NPR investigation by Katie McQue.

Malware outsourced its attack decisions to a panel of four LLMs

Malware outsourced its attack decisions to a panel of four LLMs

Cisco Talos published a new way to hunt AI-integrated malware on Tuesday, and the first thing it caught is the part worth reading twice: a Windows implant that asks four commercial language models what to do next, and then does it. Talos says CLOSEDQUORUM is the first publicly documented Windows implant to hand tactical command-and-control decisions to a panel of commercial models — DeepSeek, Qwen, Mistral and Google Gemini — and execute whatever the vote picks.