tradecraft-attack-scenarios
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where external reconnaissance findings and evidence are used to build and update an attack plan (plan.md). This creates a surface where adversarial instructions embedded in target data could influence the agent's planning strategy.
- Ingestion points: Reconnaissance data (recon-*) and evidence gathered from target surfaces.
- Boundary markers: No explicit delimiters or warnings are defined to separate plan logic from external data.
- Capability inventory: Creating/writing to plan.md and sequencing other tools; the skill itself does not have execution capabilities.
- Sanitization: No sanitization or filtering of external evidence is described in the methodology.
Audit Metadata