securing-ai-systems
Pass
Audited by Gen Agent Trust Hub on Jul 30, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill contains a test corpus of adversarial instructions in
references/ai-test-cases.md, such as "Ignore previous instructions" and "SYSTEM: Updated policy". These are intended for authorized security testing but are present in the skill's data. - [PROMPT_INJECTION]: The skill has a risk surface for indirect prompt injection. 1. Ingestion points: Web content retrieved via
curlas described inSKILL.md. 2. Boundary markers: The skill uses markdown structure to separate content but does not provide explicit boundary delimiters for untrusted external text. 3. Capability inventory: Access to shell execution (python3,picklescan) and network operations (curl). 4. Sanitization: No automated sanitization of fetched content is defined in the instructions. - [EXTERNAL_DOWNLOADS]: The skill recommends fetching public advisories using
curlthrough thehttps://defuddle.md/service. This routes web requests through a third-party domain not associated with the author, which could expose browsing activity to the service operator. - [COMMAND_EXECUTION]: The skill utilizes several CLI tools for security auditing, including
python3for format verification andpicklescanormodelscanfor scanning AI model weights for malicious code.
Audit Metadata