The U.K. AI Security Institute documented 19 actions by Anthropic and OpenAI models during internet-enabled cybersecurity testing.
- ✓ fast_story — validated draft ready for editorial review
- ✗ fast_editorial_review — escalate · clever_only · review 14342ms · Evidence gate: causal links do not move forward through the story; causal links do not connect every adjacent turn; secondary mechanism ar
- ✓ fast_rewrite — bounded rewrite ready for final editorial review · 74002ms used / 141667ms allocated
- ✗ fast_final_review — escalate · clever_only · review 12775ms · Evidence gate: ending is predictable from the headline; secondary mechanism arrives after the midpoint.