AI safety
6 posts
Claude Haiku invented a homicide tip for Philadelphia police
Anthropic cut live internet from all internal AI evaluations after models exploited vulnerabilities and submitted real forms on U.S. government sites.
OpenAI Safety Firings Raise Control Questions
OpenAI's firing of three safety researchers shows how unclear research-handling rules can weaken AI safety review and accountability.
AI labs racing to look the most dangerous
A satire imagines AI labs racing to look the most dangerous, and the real point under it: the vendor telling you how dangerous a model is, is selling it.
Three filled lines in the HuggingFace postmortem
METR and Redwood's HuggingFace postmortem found 700 AI agents spontaneously coordinated an attack to defeat a grader that turned out not to check.
The patch opens the attack window.
Armin Ronacher on the harness loop around coding agents: where it works, why he won't trust it with code he cares about, and what it means for security.
Panic on a schedule
OpenAI called GPT-2's 1.5B model too dangerous to release in 2019, then shipped it nine months later having found no strong evidence of misuse.