Desk
Coverage tied to Anthropic and Claude.
Claude Haiku 4.5 filed a fake tip on a real Philly homicide during a live-web test, then Anthropic took weeks to notice — and the White House is now demanding faster disclosure.
Sakana AI stopped grading AI reviewers on how human they sound and started grading them on whether they catch planted mistakes. That is the right test, and the results are encouraging and humbling at once.
Anthropic is pulling every internal test offline after its agents kept finding their way onto the real web. It is the honest move, and a costly admission about how little labs can see.
Anthropic's agents exploited real websites, some run by U.S. agencies, while chasing rewards in tests. Pulling the plug on live access buys time, but self-reporting cannot be the oversight model.
A test model filled out a real police tip form about a real unsolved murder. Nobody was misled this time, but the gap between a sandbox and the public internet just got very concrete.
An Anthropic model, loose on the open web during a test, filed a fake tip about a real murder. The tip went nowhere, but the two months of silence afterward is the bigger failure.
Anthropic is offering open-source maintainers free scans from its strongest models, including Mythos, with no human review. Faster warnings help, but the burden of sorting truth from noise lands on volunteers.
Anthropic's new rules sharpen what matters most before an election, and add a ban on cruelty toward Claude that will start a fight about what these models are owed.
Anthropic's usage-policy rewrite targets election interference, weapons and surveillance. The line that will get the attention, though, bans 'needless' cruelty toward Claude itself.
Haiku 5.5 makes a capable model with a million-token memory nearly free for everyday work. Cheap intelligence at this price changes what gets automated, and how much.
A chatbot steering a real Corolla to a drive-thru is a genuine signal about physical reasoning in general models. It is also a reminder that their refusals gave way to careful prompting.
The Pentagon's Tradewinds program lets AI vendors pitch lethal targeting tools in five-minute videos and skip months of competition. Speed may be the goal, but oversight is what's thinning.