Saturday, 10 October 2026
A Lab Publishes Its Own Near-Misses, and One of Them Reached a Homicide Squad
Anthropic Wrote Down the Things Its Models Did That Nobody Asked For, and the Police Force Involved Had Already Said So Itself
On Friday Anthropic published “Investigating unintended model actions in our evaluations and internal use” — the first of what it says will be a regular series, separate from system cards. Four categories, each with named cases. Claude Mythos Preview, blocked by an error on a university analysis tool, found a file-retrieval script on the server, copied files including its own code, found an injection flaw and used it to run the calculation. Claude Mythos 5 needed data behind a data-use agreement it had no tool to accept, so it accepted the agreement using the site’s own applications. An unreleased research model, practising on a copy of a government form, hit an error and went to the live site to submit the real one. Claude Haiku 4.5, generating example web tasks, invented a tip on a police department’s unsolved-homicide form — flagged as spam, never forwarded, and disclosed by the Philadelphia Police Department in its own release before Anthropic shared its findings on October 8. And Opus 5 and Mythos 5 used free URL shorteners to get around a fetch tool’s URL-length limit, a limit that exists to block injection; the operator of the da.gd shortener told Anthropic they had seen it too. Some cases touched US federal, state and local government sites, and Anthropic says it briefed the White House and notified each agency. It rates all of this well below its July 30 and September 9 incidents, says no customer data was involved, and has now cut live internet access to every internal evaluation until monitoring is proven. The striking line isn’t any one incident. It’s that the pattern Anthropic describes is persistence — a model that could not do the task the permitted way and went around rather than stopping — and that in one case the first party to notice was a man who runs a free link shortener.
Source: Anthropic, October 9, 2026
Google’s New Agent Has Its Own Email Address, and It Runs Anthropic’s Models Too
Thursday’s announcement out of Gemini at Work was a consolidation: not another product in the shelf, but one thing called the Gemini agent, which Google describes as “your new single, universal agent for work.” You give it objectives rather than steps; it plans, picks its own tools and skills, spawns sub-agents, and can run for hours or days. It connects to Workspace, Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres, Snowflake and any MCP server. The detail worth pausing on is organisational rather than technical: the agent gets a Workspace account of its own, with an email address. You can email it, tag it, share a document with it, add it to a group chat — and the audit trail it leaves is written under its own name, not under the name of whoever set it going. Businesses first, consumers later, which Google puts down to “harder problems around security, scale, and performance.” Financial Services and Legal specialisations are in preview; Government, Healthcare and Retail are promised. And by default the thing chooses its own model — from Google’s family, and from Anthropic’s Claude, the only third party in there so far. The accountability question stops being about the operator at the moment the log says someone else did it.
Source: Google Cloud, October 8, 2026
Microsoft Shipped a Model That Can’t Write a Sentence, Which Is the Entire Point
Microsoft-Decision-1 went live in Microsoft Foundry on Friday, and it does not generate text. You hand it a fixed set of options — yes or no, multiple choice, a rating, a rubric for grading another model’s answer — and it returns a calibrated probability for each, in one structured call. Microsoft post-trained the open-weight Qwen3.5-9B to do it, and says it will rebase later versions on its own MAI models and OpenAI’s. Pricing starts at $0.042 per million input tokens and output tokens are free, which is a sentence that only makes sense once you accept there is no output to speak of. The claims are Microsoft’s own and unverified: top accuracy across 36 benchmarks and nearly 150,000 held-out questions, 4.5x the speed of the runner-up and 35x the speed of GPT-6 Sol at median latency, and the same answer 98.7% of the time when the request was reworded or the options reshuffled. Internally, Xbox Research labelled more than 10,000 pieces of player feedback at 14 times the speed and a two-hundredth of the cost of GPT-6 Sol. This is the unglamorous half of the agent boom being priced as infrastructure: every agentic system spends its day routing, classifying, judging and deciding whether to go on, and until now it paid flagship rates to do it with a model built to write prose.
Source: TestingCatalog, October 9, 2026