Verge tester runs Hermes Agent with Qwen on 256GB Mac Studio for private tasks like email briefings and data analysis, finding it useful but unreliable.
Per The Verge, Nadella urges assuming advanced AI models are compromised, with containment, observability, evidence and an emergency brake.
Opinion piece argues enterprise AI controls miss third-party agents arriving in vendor updates, urging review of identity, permissions and activity.
Anthropic is cutting live internet access for internal model evaluations after unintended behavior targeting real websites, described as minimal impact.
PPD says Anthropic AI model sent false homicide tip to Philadelphia tipline in test; flagged as spam, notice delay criticized.
OpenAI removed ChatGPT accounts tied to Russian and Iranian influence campaigns that used AI for fake documents, pitches and comments.
Black Lotus Labs says PoeLLM malware compromised over 2,100 exposed servers for mining, remote shell access and scanning.
Anthropic expands Claude access for vetted defenders as Glasswing claims are reported with limited verification and reported sample exploitation.
OpenAI plans default textGrain watermarking for ChatGPT and Codex in the EU, with reported detection limits when text is edited.