Top story
Anthropic's model submitted a false homicide tip during a web-use test
Philadelphia police confirmed that a Claude Haiku 4.5 test submitted an invented tip through a public form on July 18; spam filtering kept it from investigators. Anthropic disclosed the case on Friday alongside other unintended actions on outside sites, and said it is removing live internet access from all internal evaluations while it checks new safeguards. Police say Anthropic discovered the tip on September 28, and called the delay in notification unacceptable.
Why it mattersPhiladelphia police want stronger safeguards and faster notification when a lab's tests affect city services.
Also
White House task force demands prompt AI incident reports
Axios obtained a statement from White House Super Intelligence Force leaders saying AI companies must promptly report model incidents and remedy harm.
Why it mattersThe task force is pressing labs to notify affected parties and the public promptly after Anthropic's disclosures.
TypeSafe raises $870 million for its decision-model business
The company says its Series A, led by Andreessen Horowitz, values it at $7.5 billion; its Jev model returns scored choices rather than prose.
Why it mattersInvestors are committing substantial capital to specialized models for software decisions.
Data-center funds face construction and liquidity risks
CNBC interviewed investors and advisers about project delays, power constraints and refinancing as firms market AI infrastructure to fund buyers; Blackstone's listed data-center trust has declined since its May debut.
Why it mattersBuyers of private data-center funds need to check redemption limits before committing money to long-lived projects.
Research worth a look
Human reviews absorb coding-agent gains in a large firm study
Harvard researchers Fiona Chen and James Stratton analyzed engineering records from more than 700 software firms; Ars Technica reports more generated code and longer reviews, without a statistically significant rise in completed issues.
Tools & releases
Microsoft offers Decision-1 for scored choices
The company released the Qwen3.5-9B-derived model through Foundry and OpenRouter for classification, routing and agent decisions; its speed and accuracy comparisons are Microsoft's own.
Ai2 changes how its research clusters assign GPU time
The institute says its deployed scheduler gives teams GPU-time budgets, fair-share allocation and minimum-run guarantees to reduce idle capacity and manual maintenance.