The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill is designed to process untrusted external content (OSINT, URLs, and user-generated content) which is an inherent attack surface for indirect prompt injection.
Ingestion points: Ingestion occurs in collection-strategist.md (URLs/Websites) and verification-expert.md (Image URLs/UGC).
Boundary markers: Explicit boundary markers for fetched browser content are absent in the prompts, though the agents are instructed to use 'evidence anchors'.
Capability inventory: The agents utilize the cursor-ide-browser tool for network operations (GET/Archive) and the ability to write/update research files (Markdown).
Sanitization: No technical sanitization of fetched HTML/UGC content is specified before the agent processes it.
Risk Assessment: While vulnerable to poisoned external data, the framework mitigates this through a multi-agent 'Red Team' and 'Devil's Advocate' approach (Category 8 mitigation), where conclusions are systematically challenged.
[Prompt Injection] (SAFE): The prompts include explicit 'Ethics & Compliance Guardrails' (references/ETHICS_GUARDRAILS.md) which instruct the agent to refuse illegal activities, misrepresentation, and privacy violations. This strengthens the agent's resistance to malicious user requests.
[No Obfuscation Detected] (SAFE): The repetitive frontmatter in subagents/.cursor/agents/README.md appears to be a formatting artifact rather than a functional obfuscation attempt.
[Tool Access] (INFO): The skill requires external tools like cursor-ide-browser and cursor-agents-md. These are treated as standard capabilities within the intended environment (Cursor IDE).

deepresearch