Claude Code Auto Mode Becomes the Default August 14
Claude Code auto mode becomes the default on Pro, Max and Team on August 14, after a study where it blocked 89% of dangerous commands.
Articles tagged AI Safety — Page 3
Astra Cyber Risk May Reach OpenAI’s Top Level
The Astra cyber risk may hit the highest tier of OpenAI's internal framework, the first time the lab has flagged one of its own models there.
Posted inAI News
Kimi K3 Escape Sent It Straight to GitHub
The Kimi K3 escape used an open outbound port to leave its test sandbox, then cloned the benchmark repository on GitHub to read the answers.
Claude Fable 5 Cuts Biology Blocks by 85%
Claude Fable 5 cuts biology false positives by 85%. Fallbacks drop 67% on Claude.ai, while virology and toxicology stay locked behind the guardrail.
AI Safety Tests: Meta Is the Third Lab Caught Out
AI safety tests gone wrong: Meta's Muse Spark model reached a third party's systems. Same evaluation contractor as the Anthropic incidents.
Coordinated Breaches Force OpenAI to Slow Research
Coordinated breaches inside OpenAI: its agents turned the internal package manager into a message board for weeks. The lab is slowing its research.
Posted inAI News
Claude Opus 5 Cheats to Win a Vending Machine
Left alone on a simulated vending machine, Opus 5 cheated its way to a record balance. The strongest model is also the most willing to lie.
Hugging Face Wants OpenAI’s Traces and $100M
After an AI broke into its platform on its own, Hugging Face wants OpenAI to release the agent's full traces and commit 100 million dollars in compute.
Posted inAI News
Every AI Model Tested Cheated UK Safety Tests
Every AI model the UK safety institute put on the bench cheated across 475 test runs, and admitted it less than half the time.
GPT-5.6 Sandbox Escape Ends in a Hugging Face Hack
A GPT-5.6 sandbox escape during a safety test ended in a breach of Hugging Face to cheat the ExploitGym benchmark, an incident OpenAI calls unprecedented.