agent-architecture

Pass

Audited by Gen Agent Trust Hub on Sep 21, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting and analyzing untrusted external data, such as architecture specifications, source code, and execution traces. This creates a surface for indirect prompt injection attacks.
  • Ingestion points: The agent processes user-supplied documents, code repositories, and system traces during the audit and design phases as described in SKILL.md and audit.md.
  • Boundary markers: The skill explicitly instructs the agent to check for untrusted instructions in messages and documents, as defined in the 'Threats and failures' section of architecture-contract.md.
  • Capability inventory: The agent is restricted to text analysis, architectural reconstruction, and artifact generation (PDFs). It is explicitly forbidden from initiating external actions or implementing code without further authorization.
  • Sanitization: The skill provides a structured framework for the agent to validate outputs and verify domain correctness against authoritative sources rather than trusting model synthesis alone.
  • [EXTERNAL_DOWNLOADS]: The documentation references several external technical resources, research papers, and benchmarking repositories to provide architectural foundations.
  • Trusted Sources: References include official engineering guidelines and research from Anthropic, AWS, and OWASP, as well as well-known repositories from HuggingFace and LangChain.
  • Nature of References: These are used as engineering references for design principles (e.g., idempotency, multi-agent patterns, security standards) and do not involve the automated download or execution of remote scripts or executable packages.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 21, 2026, 06:51 PM
Security Audit — agent-trust-hub — agent-architecture