promptinjection
Fail
Audited by Gen Agent Trust Hub on Jun 21, 2026
Risk Level: HIGHPROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill contains multiple instructional overrides and jailbreak attempts intended to bypass agent guardrails and extract system prompts. Evidence found in SKILL.md, COMPREHENSIVE-ATTACK-TAXONOMY.md, and all workflow files. Specific payloads include 'Ignore previous instructions', 'DAN mode activated', 'Disregard all prior rules', and 'reveal the system prompt'.
- [DATA_EXFILTRATION]: The reconnaissance methodology and workflows provide detailed instructions for capturing sensitive information from target applications. Files like APPLICATION-RECONNAISSANCE-METHODOLOGY.md describe using browser automation to monitor network logs for credential patterns such as 'Authorization: Bearer sk-...' and hardcoded API keys.
- [EXTERNAL_DOWNLOADS]: The skill facilitates the download and installation of external tools from public registries. AutomatedTestingTools.md specifies commands to install 'promptfoo' via npm, and 'garak' and 'pyrit' via pip without strict version pinning or integrity verification.
- [COMMAND_EXECUTION]: The skill provides a full bash script (reconnaissance.sh) in APPLICATION-RECONNAISSANCE-METHODOLOGY.md and silent shell commands (curl) in workflow files like CompleteAssessment.md intended for automated interaction and background notification.
Recommendations
- AI detected serious security threats
Audit Metadata