brida-reflex-software-issue-triage

Pass

Audited by Gen Agent Trust Hub on Sep 21, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of software issue reports, creating a surface for indirect prompt injection.
  • Ingestion points: Software issue titles, bodies, and repository context are ingested as state for the software-issue-triage Reflex, as defined in references/custom-reflex.json and SKILL.md.
  • Boundary markers: The documentation in references/playbook.md suggests using "bounded body/summary" and "deterministic signals," but does not define specific escaping or delimiter patterns to isolate the untrusted issue body from the agent's instructions during semantic analysis.
  • Capability inventory: While the Reflex itself is described as recommendation-only, the skill instructs the agent to choose an integration surface (MCP, SDK, or REST) to implement the use case, which provides a path for actions to be taken based on analyzed content.
  • Sanitization: The skill delegates parsing and initial validation to "deterministic host code," but lacks specific instructions for the agent to sanitize or escape the semantic content before LLM processing.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 21, 2026, 08:22 PM
Security Audit — agent-trust-hub — brida-reflex-software-issue-triage