Secure self-modifying agent systems

Audit agent and LLM system for injection, prompt-leak, and data-exfiltration risks

Production agent deployment where adversarial input and prompt injection are realistic threats.

• · agent code
• · prompts
• · data handling flow

• · risk report
• · remediation steps

Code and prompts available; data classification known; threat model defined

False negatives miss novel injection patterns; false positives block legitimate uses