safety
2 articles
Anthropic Analyzed Millions of Claude Code Sessions — Your Agent Can Handle Way More Than You Let It
Anthropic's Claude Code AI agent study: autonomous runs doubled (45+ min), experienced users auto-approve 40%+ sessions. Claude clarifies more than interrupted. 73% of API actions still human-in-loop. Key: models handle more autonomy than users grant ('deployment overhang').
AGENTS.md Can't Stop a Rogue AI: jzOcb's 4-Layer Defense System
After letting an AI agent manage a server and hitting 7 disasters in one day, the lesson: use code hooks instead of markdown rules, build a 4-layer defense system