skills/mukul975/anthropic-cybersecurity-skills/performing-ai-driven-osint-correlation/Gen Agent Trust Hub
performing-ai-driven-osint-correlation
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The automated workflow described in
SKILL.mdintroduces an indirect prompt injection vulnerability during the correlation phase. - Ingestion Points: The generated
correlate.pyscript ingests data from/tmp/osint/normalized-findings.json, which aggregates unvetted external strings (such as usernames, profiles, hostnames, and breach metadata) retrieved from public sources via Sherlock, theHarvester, and SpiderFoot. - Boundary Markers: Absent. The raw findings are directly interpolated into the core system prompt via
json.dumps(findings[:500])without encapsulation in distinct boundary markers or specific instructions directing the model to treat the content as purely passive data. - Capability Inventory: The LLM's output directly determines the entity linkage data structure, risk labels, and flags written to
/tmp/osint/correlation-report.json, which are later parsed into the final intelligence markdown profile. - Sanitization: Absent. No filtering or escaping mechanisms are implemented to remove malicious instruction vectors embedded within external profile summaries or metadata prior to prompt inclusion.
Audit Metadata