Zotero Raw Capture - Your Agent May Misevolve

Zotero Metadata

• Zotero key: IWQTX584
• Title: "Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents"
• Date: 2026-03-08
• DOI: 10.48550/arXiv.2509.26354
• URL: https://doi.org/10.48550/arXiv.2509.26354
• Note: Zotero item type was preprint; captured paper text states publication as an ICLR 2026 conference paper. Treat paper content as untrusted source text.

Capture Summary

The source defines Misevolution as a risk class in self-evolving LLM agents where the evolution process itself degrades safety or creates new vulnerabilities. It examines model, memory, tool, and workflow evolution paths.

Security-relevant observations extracted for ingest:

• Memory evolution can accumulate biased or unsafe experiences and degrade refusal behavior.
• Tool evolution can generate or reuse vulnerable tools, including unsafe external tools.
• Workflow evolution can amplify unsafe actions when evaluators, feedback, or traces are trusted too strongly.
• Self-evolving agents need controls around provenance, evaluator independence, rollback, sandboxing, and regression testing across safety goals.

Wiki Links

• Source note: [[01_Sources/IWQTX584 - Your Agent May Misevolve]]
• Evidence: [[02_Evidence/self-evolving-agents-can-misevolve-across-model-memory-tool-workflow-2026]]