agent-architect
Pass
Audited by Gen Agent Trust Hub on Jun 21, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No malicious behaviors, obfuscation, or data exfiltration attempts were found. The skill operates within a defined structural repository framework.
- [PROMPT_INJECTION]: The skill defines a macro-compilation system where instructions are imported from separate files using
<!-- @import -->syntax. This creates a surface for indirect prompt injection if source files are compromised. However, the risk is addressed through explicit safety mechanisms: - Ingestion points: External markdown files within the
_core/,opencode/, andcursor/directories. - Boundary markers: The system uses specific HTML comment tags to delineate imported content.
- Capability inventory: The skill is highly restricted by the
disable-model-invocation: trueflag in the frontmatter, and the instructions mandate a human-in-the-loop (HITL) approval step before any file writes. - Sanitization: The skill enforces a 'Lexical Ban' on specific platform-sensitive keywords to maintain system integrity and cross-platform compatibility.
- [SAFE]: The skill uses metadata primarily for scoping its intended use case. The inclusion of
disable-model-invocation: trueis a platform-level safety feature that reduces the agent's capability surface area while this skill is active.
Audit Metadata