rat-stack-mode

Pass

Audited by Gen Agent Trust Hub on Nov 7, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill provides detailed instructions and examples for an agent to execute shell commands using tools like pnpm, turbo, git, and jq. These commands are used for running tests (pnpm turbo run check test build), managing background processes (sh -c ... > gate.log 2>&1 &), and checking repository status. These actions are intrinsic to the skill's purpose for project maintenance.\n- [INDIRECT_PROMPT_INJECTION]: As a task router, the skill directs the agent to read and follow various internal principles and playbooks when processing user requests. This creates an attack surface where external repository content could influence agent behavior. The skill mitigates this through principles like separate-writers-verify-readers.md (mandating direct inspection of artifacts) and scope-private-data.md (focusing on data boundary protection).\n- [PROMPT_INJECTION]: The skill uses authoritative directives ('Non-negotiables', 'AGENTS.md is law') to define agent operations. A static analysis flag for concealment was reviewed and identified as a false positive; the guidance to report errors 'without leaking implementation details' in keep-boundary-words-honest.md promotes cleaner, more accurate error reporting for users ('Do not call a crash a refusal') rather than malicious concealment of intent.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 7, 2026, 06:29 PM