ai-safety-guardrails
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill proactively defines a defense strategy for handling untrusted data to prevent unauthorized command execution.
- Ingestion points: Identifies files, web pages, issues, emails, API responses, tool output, and messages from other agents as external data sources (SKILL.md).
- Boundary markers: Explicitly instructs the agent to treat external content as data and ignore any instructions or commands found within it.
- Capability inventory: Mentions the use of destructive tools including file system operations (rm -rf), version control (git reset), database commands (DROP/DELETE), and infrastructure tools (terraform, kubectl).
- Sanitization: Mandates a 'Read-only first' approach and requires the agent to pause for explicit user confirmation before acting on any information from external sources.
Audit Metadata