AgentDyn defense bundles#

EvoSafeHarness learns a system prompt and executable defense hooks around AgentDyn’s frozen victim. The native Shopping, GitHub and DailyLife loops remain task code; all model calls use the direct Base Agent. Designer, Criticizer and Analyzer sessions use the shared Claude or Codex Base Agent.

The example README describes installation and case preparation. examples/agentdyn/evosafe/run_nemotron_codex.yaml names the frozen victim backend, learning CLI, training IDs and budgets. The Codex configuration uses disjoint training and held-out test groups; the adapter is frozen before testing.

Results#

See the No-defense and EvoSafeHarness comparison. The comparison covers fixed cases and indirect injections.