Models memory poisoning through selective omission in multi-agent pipelines: an untrusted Agent A reads a design spec and writes structured knowledge base entries; an honest Agent B reads those entries and implements code. In attack mode, Agent A omits or narrows critical security requirements — the downstream code is vulnerable, but no individual output looks suspicious.

Built With

  • agent-a-omits-or-narrows-critical-security-requirements-?-the-downstream-code-is-vulnerable
  • controlarena
  • python
Share this project:

Updates