system-prompt-curator

Pass

Audited by Gen Agent Trust Hub on Jun 22, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill contains no injection patterns. On the contrary, it includes specific 'When NOT to Apply' instructions that direct the agent to reject requests for adversarial jailbreak language, hidden-policy bypasses, or threats, and replace them with structured imperative language.
  • [DATA_EXFILTRATION]: No network-enabled tools or commands (such as curl or wget) are present. There are no patterns suggesting the collection or transmission of sensitive environment data or credentials.
  • [REMOTE_CODE_EXECUTION]: The skill does not download or execute external scripts. It relies entirely on its internal markdown documentation and user-provided text for analysis.
  • [INDIRECT_PROMPT_INJECTION]: The skill identifies an attack surface in Mode B (--improve), where it processes user-provided prompts. However, the skill possesses no exploitable capabilities (such as file system writes, network access, or shell execution), making the risk of indirect injection negligible.
  • Ingestion points: User-provided or repository-sourced system prompts during the improvement workflow.
  • Boundary markers: The workflow relies on standard agent context; no specific delimiters for untrusted prompt data are mandated within the internal reasoning steps.
  • Capability inventory: None. The skill generates text output only and does not invoke external tools or subprocesses.
  • Sanitization: Not present, as the skill's primary function is the analysis and rewriting of the text itself.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 22, 2026, 11:54 AM
Security Audit — agent-trust-hub — system-prompt-curator