The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple external sources to generate vulnerability reports and executable Python code, creating a surface for injection attacks.
Ingestion points: Workflow Step 2 in SKILL.md and Section 1 in reference/poc-methodology.md instruct the agent to deep-dive into vendor advisories, GitHub Security Advisories, Exploit-DB, technical blogs, and conference talks.
Boundary markers: The skill provides high-level rules (e.g., "Least harm", "Safety constraints") but lacks structural delimiters or explicit instructions to ignore embedded malicious prompts within the fetched advisory text.
Capability inventory: The skill generates and writes files (poc.py and report.md) to the output directory.
Sanitization: There is no evidence of content filtering or sanitization applied to the external data before it is interpolated into the prompts used for script and report generation.
[DYNAMIC_EXECUTION]: The primary function of the skill is generating Python scripts (poc.py) based on external descriptions of vulnerabilities. If the source data is poisoned, the generated script could perform unintended actions despite the safety guidelines provided in the methodology.
[EXTERNAL_DOWNLOADS]: The skill queries external security databases and repositories for research purposes.
Evidence: It fetches data from services.nvd.nist.gov, github.com/advisories, exploit-db.com, and cdn2.qualys.com. These are recognized services in the vulnerability research domain.

cve-poc-generator