qa-persona-testing
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill operates by automating a browser to visit external websites and analyze their content, which introduces a vulnerability to instructions embedded in those pages.
- Ingestion points: Browser accessibility tree snapshots and screenshots retrieved via tools like
browser_snapshotandtake_snapshotas described inreferences/browser-execution.md. - Boundary markers: The skill employs a dual-channel logging system (in-character persona voice vs. out-of-character observation) in
assets/session-log.mdto help the agent distinguish between simulated reactions and QA analysis, but it does not provide explicit technical delimiters or instructions to ignore malicious commands within the scraped UI data. - Capability inventory: The agent has significant capabilities including browser navigation, interaction (click, type, fill form), and the ability to execute generated shell scripts.
- Sanitization: There is no mention of sanitizing or filtering the DOM or accessibility tree data before it is processed by the agent's logic.
- [DYNAMIC_EXECUTION]: The skill instructions involve the creation and execution of code at runtime for specific testing scenarios.
- Evidence: In
references/browser-execution.md, the skill directs the agent to "Generate a Playwright script per scenario that replays the persona's known path" as a fallback for CI/CD environments or when MCP tools are unavailable. - Risk: While used for legitimate regression testing, the dynamic generation and execution of scripts (likely via a shell) is a form of runtime code assembly that could be abused if the generation process is influenced by malicious input.
Audit Metadata