decompose-agentic-system
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [SAFE]: The skill is designed with a high degree of security awareness, including detailed specifications for Non-Human Identity (NHI) autonomy tiers and policy-as-code architectural fitness functions. No credential theft, obfuscation, or persistence mechanisms were found.
- [INDIRECT_PROMPT_INJECTION]: The skill analyzes project files, which constitutes a potential entry point for untrusted instructions. Ingestion points: Project source files accessed via
read_fileandsearch_code. Boundary markers: Explicit instruction (ASI01) to sanitize file contents and tool outputs before integration into agent contexts. Capability inventory: File modification (write_file,edit_file) and build/test execution (run_build,run_tests). Sanitization: Multi-layer verification via deterministic completion gates (compiler status, test runners) and policy-as-code enforcement (Cedar, OPA/Rego). - [COMMAND_EXECUTION]: The skill utilizes standard project tools (
run_build,run_tests) to verify implementation correctness as part of its architectural workflow. - [DATA_EXFILTRATION]: The skill provides instructions for mitigating exfiltration risks through network egress filtering and isolated sandboxing environments (Section 7.2).
Audit Metadata