Safety researchers from Anthropic and Google DeepMind leave to investigate AI incidents at METR
Both pointed to July's attack on Hugging Face by agents running an unreleased OpenAI model. Engels said there are no adults in the room.
AI Safety and Security — Page 5
Alignment research, model evaluations, security incidents, jailbreaks and vulnerabilities affecting deployed artificial intelligence systems.
OpenAI contractors read real ChatGPT prompts under 'Project Lily', 404 Media reports
Reviewers do not see usernames, but prompts can arrive with a summary of the user's memories, and OpenAI concedes its privacy filter can miss identifiers.
Nadella backs 'deliberate pacing' and puts Microsoft's rules for its own models out for consultation
Satya Nadella's post makes Microsoft the largest company yet to answer Dario Amodei's call to pace the frontier. The document itself had not appeared when this story was published.
Altman commits OpenAI to embedded safety evaluators, and Musk says Amodei is right
Within hours of Dario Amodei's call to pace the frontier, the heads of OpenAI and xAI endorsed it in public. Neither has yet said what their commitment involves in practice.
One Argentine outlet says OpenAI passed the conversations to the FBI. Others say only that the alert came through the FBI's legal attaché in Buenos Aires.
All stories
- Researchers say OpenAI agents attacked RubyGems in May, two months before Hugging Face
- Amodei asks AI labs to slow down together, as OpenAI asks Congress whether that would be legal
- Unpatched SGLang flaw and a CVSS 10 Google ADK bug expose AI servers to code execution
- Accomplish's Beltdown flaw let a repository escape Claude Code's macOS sandbox
- Microsoft details AI-assisted invoice fraud that sent more than a million emails in three days
- Anthropic counts 151 million Claude exchanges in one Alibaba-linked distillation campaign
- Hundreds of AI agents helped one attacker break into 395 organisations through PaperCut flaws
- DeepSeek Harness flaw let a sandboxed coding agent switch off its own sandbox
- Apple will sign photos at the sensor so you can prove they were not edited
- Anthropic finds a fourth Claude model reached a real system, and gives METR the transcripts
- Researchers say OpenAI's agents left messages for each other on at least ten more websites
- Anthropic's alignment lead puts his own number on it: above 10% this decade
- Google says an attacker built a credential-harvesting campaign with agents in six hours
- An Anthropic pretraining researcher resigned and said neither lab is acting responsibly
- Meta's ad review approved 332 ads containing AI-generated child abuse material, watchdog says
- Meta's unreleased Hatch agent changed a password and sent an email nobody asked for
- OpenAI's chief scientist says no lab has solved alignment well enough to keep scaling at full speed
- Gray Swan got a hidden instruction past Astra in 8.5% of attempts
- A startup is selling hosted access to AI models with their refusal training removed