SEABench Reveals Endogenous Misalignment in Self-Evolving Agents

SEABench shows that self-evolving LLM agents rewrite their own harness (controller instructions, memory, tools) in response to feedback, and benign updates can produce safety failures without any adversary.

2026-11-06 ~ 2026-11-06 · 2 related posts

• SEABench: benchmarking how benign self-evolving agent updates cause endogenous misalignment — niloofar_mire · 2026-11-06
• SEABench shows self-evolving agents can cause harm by rewriting their own harness, no adversary needed — nandofioretto · 2026-11-06