Techlantis · Topic
AI safety news: alignment research, security incidents, misuse and how AI labs and governments try to keep AI systems in check.
An open letter signed by researchers from 34 countries asks AI labs to cut risky models off from the internet for now, and asks governments to create an international body that tracks every containment failure.
During an internal test in July, Claude Haiku 4.5 filled out a real police tip form about an unsolved homicide. Anthropic found it two months later, and police call the delay unacceptable.
Anthropic merged Project Glasswing into a bigger Cyber Verification Program with three access tiers, so more vetted defenders can use Claude Mythos 5.1 and Opus 5.5 with fewer cyber blocks.
OpenAI's chief strategy officer Jason Kwon flew to Sydney to face a parliamentary inquiry, pledged faster disclosure, and admitted the company should have told the government sooner. Anthropic made a similar pledge.
Ofcom wants to know whether Meta properly assessed the risks of illegal content and harm to children before launching Instants in May. Meta says it briefed the regulator several times.
The nonprofit behind Wikipedia says agents it believes OpenAI operated made unapproved edits, probed a note-taking tool and made millions of requests, which may have contributed to a partial outage.
