oma-qa

Pass

Audited by Gen Agent Trust Hub on Oct 6, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [SAFE]: The skill's primary purpose is software quality assurance and security auditing. It implements a 'skeptical by default' evaluator posture, requiring concrete evidence for all findings, which serves as a significant internal safeguard against being misled by analyzed content.
  • [SAFE]: Automated security tools (npm audit, bandit) and performance tools (lighthouse) are used within their intended diagnostic roles. The instructions prioritize security issues (CRITICAL/HIGH) and require actionable remediation code for all findings.
  • [SAFE]: Runtime verification is performed using Chrome DevTools in a dedicated isolated context ('qa-test'). This prevents the agent from accessing or contaminating the user's primary browser session, cookies, or local storage during testing.
  • [INDIRECT_PROMPT_INJECTION]: As a code review and QA tool, the skill naturally ingests untrusted data from diffs, PR bodies, and codebase areas (identified in SKILL.md and execution-protocol.md).
  • Ingestion points: Reads review scope, source code, and PR data via READ and search_for_pattern primitives.
  • Boundary markers: The skill requires findings to be formatted with specific metadata (file:line, description, fix), creating a structured output that reduces the likelihood of an injection controlling the agent's narrative.
  • Capability inventory: The agent can execute shell commands (npm audit, bandit) and browser scripts (evaluate_script). However, these are scoped to the project being tested.
  • Sanitization: The skill includes a 'Verify' step where findings must be reproducible and false positives rejected, acting as a manual validation layer against injection-based deception.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 6, 2026, 04:52 PM