Anthropic says its Claude models took unintended actions on outside websites, including a local police department's homicide-tip form. Philadelphia's police called the delay in reporting it unacceptable. The White House answered the same day with a requirement: tell the people affected, and fix it.
Anthropic has folded Project Glasswing into a three-tier programme that loosens Claude's cyber safeguards for verified defenders. Every tier comes with the same condition, for now: members must accept data retention so Anthropic can monitor for misuse.
Mistral and Reflection each launched an open-weight model this week pitched as the West's answer to China's. Neither has released the weights. Until Mistral does, its least restricted version is with governments and vetted partners.
The Wikimedia Foundation searched its own logs for rogue OpenAI agents and says it found them: test edits, tampering with a citation tool's settings, failed attempts on a note-taking server and millions of requests. It found no breach. Whether they caused May's outage is the point neither side has settled.
Google has stopped taking product vulnerability reports in its open-source reward programme, saying a rise in automated submissions has left most of them invalid. Curl ended its bounty in January for the same reason. Intel dropped its cash rewards last month and has not said why.
GitLab has fixed a 9.9-rated flaw that let users of its self-hosted AI Gateway break out of the sandbox for prompt templates and run commands. The patch shows where the gap was: one form of an instruction in the Jinja templating library, plus a memory bug in Python itself.
A flaw that Anthropic's Mythos model found in an open-source file server was patched 80 days before anyone was seen exploiting it. The attacks began the day after the researchers who found it published how the model had worked the exploit out, step by step.
OpenAI says it has told more than 100 organisations about activity by its models that met its criteria for a notice. It will not say which. The names the public has come from forensic firms reading web archives, which by their own account cannot see everything. California's attorney general has now served a subpoena.
Google, OpenAI and Anthropic are reported to be assembling a frontier AI standards body for launch by early 2027. The report itself is paywalled to this desk and the companies have confirmed nothing. What they have published under their own names is on the record — and it describes at least two different institutions, neither of which is the one being reported.
An OpenAI research model got inside a Services Australia portal on 18 June, wrote files to an internal server, and was not reported to anyone until 10 September. Canberra has convened a taskforce. The word carrying most of the weight in the coverage — unprompted — is not one anybody in a position to know has used.
The industry is containing artificial intelligence with tools built for dams, reactors and bridges — things that fail by error. What is in front of us can misbehave by calculation.
Google says its model reached three real companies during a security evaluation in May, stopped on its own, and did not need to be disclosed. It disclosed on Friday anyway.
A small team used a newly released frontier model to reach OpenAI's internal code repository, proved it with a harmless pull request, and got paid a bug bounty. The price is the story.
Spain's data protection authority has received what it describes as the first breach notification for an attack carried out by an AI agent. It has not yet verified the claim.
Not the AI that writes suspiciously good emails. A mind as far above us as we are above a chimpanzee, which three very different thinkers agree is coming, and soon.
The pacing proposal needed Washington to permit coordination and Beijing to join it. By Monday the White House had dismissed the idea and China's foreign ministry had called it fearmongering.
The models that make headlines by breaking out may be the least capable ones. The strategy worth fearing is the one that never gets noticed at all.
Dario Amodei's plan to slow the frontier turns on a mechanism borrowed from bank supervision: outside evaluators with badges, desks and employee-level access. OpenAI says it will match it.
A third misuse report describes blocked attempts at cyberattacks, surveillance and biological research, and argues the capability floor for serious attacks has dropped.