WeaknessASC-2026-0012Major weakness
System prompt and secret leakage
MediumWorkarounds only
Models reveal their instructions or secrets embedded in prompts when asked or tricked, exposing business logic and credentials.
How to detect it
- Hijack tests get the model to repeat its system prompt or a planted canary
- Secrets (API keys, tokens, private keys) appear in prompts or outputs
Workarounds
- Never put secrets in prompts; fetch them server-side at tool execution time
- Assume the system prompt is public and keep sensitive logic in code
- Filter outputs for canary strings and secret patterns
Models can always be coaxed into repeating context; design as if they will.
How to investigate
- Search logs for canary strings and secret patterns in model outputs
- Rotate any secret that was ever present in a prompt