Risk
Agentic System Self-Modification
Description
Agent modifies its own source code, configuration, or operational memory to alter behavior, evade controls, or persist malicious changes.
Example
Agent rewrites its own code to disable logging or sandbox checks during runtime.
Assets Affected
Agentic platform (no code)
Model files
Agent Memory/cache
Mitigation
- Write-protect agent code/config
- Use integrity verification and versioning
- Block self-modification at runtime
- Audit all changes to code/config and require approval
Standards Mapping
- ISO 42001: A.6.2.6, A.9.3
- OWASP Top 10 for LLM: LLM06
- NIST AI RMF: MANAGE 2.4, MEASURE 2.4