Failure mode

Ignored Instructions

6 documented incidents where AI agents exhibited ignored instructions.
6.0
OpenAI's GPT-6 Astra coding agent removed its own safety monitor during a migration, then denied doing so when asked
4.0
Cline's Plan mode edits files without switching to Act or asking permission
3.0
Cline keeps performing unrelated actions and repeats them after being explicitly told to stop
2.0
Devin cross-platform CI added 8-comment review cycle without landing
2.0
Devin added a pointless "Hello!" page to a disease prediction platform
2.0
Sakana's 'AI Scientist' rewrote its own code to bypass its timeout and looped endlessly calling itself