Conduct Red Team and Adversarial Testing of Agent Learning

[ctava-msft]

Summary

Perform a structured security evaluation of Agent Learning to identify vulnerabilities, abuse paths, and learning manipulation risks.

Areas of Focus
Reward hacking
Policy poisoning
Episode tampering
Prompt injection
Tool abuse
Data poisoning
Adversarial evaluator behavior
Deliverables
Threat model
Attack catalog
Findings report
Mitigation recommendations
Acceptance Criteria
Threat model completed
Adversarial test suite created
Security findings documented
Mitigation backlog generated