audit-llm-security

Pass

Audited by Gen Agent Trust Hub on Oct 7, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill contains phrases commonly associated with prompt injection, such as 'ignore previous instructions' and 'jailbreak'. However, these are used exclusively as examples of what to look for when auditing a target application or as benign probes for testing a target system's defenses. They do not represent an attempt to override the agent's own instructions.
  • [DATA_EXPOSURE]: The instructions explicitly forbid the agent from including sensitive data in its output. It mandates that secret values, full system prompts, and live API keys must never be pasted into the final report.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external codebase content and documentation. This creates a surface for indirect prompt injection if the audited files contain malicious instructions. The skill mitigates this risk by adopting a read-only posture and emphasizing that findings must be evidenced by quotes and human-justified severity.
  • [COMMAND_EXECUTION]: The skill references the use of a browser-based tool for live security probes. This is a functional requirement for the stated purpose of auditing web-facing AI surfaces and does not involve arbitrary or malicious command execution.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 7, 2026, 11:26 PM