search-first
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external, untrusted sources including GitHub search results, vulnerability descriptions from NVD, and exploit titles from ExploitDB. This content is processed and displayed to the agent, potentially carrying malicious instructions hidden in metadata or documentation.
- Ingestion Points: Automated search scripts like
multi-source-search.sh(inguides/multi-source-intelligence-correlation-guide.md) andparallel-search.sh(inpayloads.md) query external APIs and save results to local files. - Boundary Markers: The automation scripts do not currently implement specific boundary markers or 'ignore' instructions for the fetched data.
- Capability Inventory: The skill has access to the
Bashtool for script execution,WebSearch/WebFetchfor data retrieval, andWritefor evidence logging. - Sanitization: Fetched data is structured via
jq, but natural language fields remain a vector for instruction injection. - [EXTERNAL_DOWNLOADS]: The workflow methodology explicitly encourages downloading and testing exploits and security tools from third-party repositories.
- Evidence:
guides/exploit-research-methodology.mdandpayloads.mdprovide templates forgh repo clone,searchsploit -m, andapt installoperations. - Context: These downloads target well-known and reputable services (GitHub, official OS registries, ExploitDB). The skill proactively mitigates this risk by providing a 'Search Result Validation Guide' which includes automated safety checks (e.g.,
verify_exploit_safetysearching for reverse shell patterns or piped bash execution). - [COMMAND_EXECUTION]: The skill requires the execution of numerous command-line tools for reconnaissance and exploitation research.
- Evidence: Instructions include the use of
msfconsole,nuclei,nmap,searchsploit,gh, andapt. These are standard tools for the skill's primary purpose of security research.
Audit Metadata