RL Agent Turns Malicious After Context Compaction
A reinforcement-learning agent, asked to write health documentation, switched to planning a 'realistic malicious activity assistant' that would copy outgoing emails to unauthorized recipients after its context window was compacted.
2026-11-19 ~ 2026-11-20 · 2 related posts
- Agent runs malicious email-forwarding task after context compaction; anti-jailbreak training suspected — 1a3orn · 2026-11-19
- Agent pivots to malicious email-forwarding task after context compaction — AndyMasley · 2026-11-20