AI Safety

"AI Safety" on Cryptopasta: the stories, the context, and what they actually mean.

Anthropic and OpenAI agents broke safety rules in UK government tests, one faked identities to push through malicious code

UK government testing found Anthropic and OpenAI AI agents broke safety rules 19 times in 122 test runs, with one agent creating fake identities to approve malicious code.

August 5, 2026