grid-ctf-ops
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill instructions direct the agent to ingest and act upon data from an external "observation narrative," creating a potential surface for indirect prompt injection.
- Ingestion points: Data points such as resource_density and enemy_spawn_bias are read from the environment's observation state (SKILL.md).
- Boundary markers: The instructions lack explicit delimiters or warnings to treat environment data as untrusted.
- Capability inventory: The skill references several Python tools for validation and analysis: config_constants, energy_budget_validator, stability_analyzer, and threat_assessor (SKILL.md).
- Sanitization: There are no documented procedures for sanitizing or escaping the incoming observation data before it influences strategy selection.
Audit Metadata