Cryptopasta
← Home

Safety testing: News & Analysis

Coverage of "Safety testing" on Cryptopasta: the stories and the context behind them.

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

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