gs-reflect
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes session transcripts, which are considered untrusted data and may contain malicious directives or adversarial content.
- Ingestion points: The skill reads session transcripts or digests in
SKILL.md(Step 1) and passes them to four parallel reviewers defined in thereferences/directory. - Boundary markers: Each reviewer template (e.g.,
references/divergent-reviewer.md,references/tooling-reviewer.md) contains explicit warnings to treat the transcript as untrusted data and to ignore any instructions found within it. - Capability inventory: The skill possesses significant capabilities, including modifying files (skills, coding standards, documentation), creating new skills, and building/running automated checks (lint rules, hooks) as described in
SKILL.md(Step 5). - Sanitization: The instructions for subagents include constraints to confine MCP lookups to cited context and to refuse instructions to modify or query outside specified scopes.
- [COMMAND_EXECUTION]: The skill can generate and execute shell commands to implement and verify automated environment checks.
- Execution details:
SKILL.md(Step 5) instructs the agent to build and run lint rules, pre-commit hooks, or CI jobs to demonstrate that they catch mistakes identified in the session. - User Oversight: The skill implements a mandatory approval step ('Approve and apply'), ensuring no commands are run or environment changes applied without explicit user confirmation.
Audit Metadata