The Agent Skills Directory

[DATA_EXPOSURE]: The skill operates within a specific local workspace (~/CybosVault/private/) to store research reports, agent outputs, and logs. This appears to be the intended operational environment for the cyber•Fund investment persona. No unauthorized data access or exfiltration patterns were identified.
[INDIRECT_PROMPT_INJECTION]: The skill is highly vulnerable to indirect prompt injection as it actively ingests untrusted data from the web using MCP tools like Perplexity, Exa, and Firecrawl. It lacks explicit boundary markers (e.g., delimiters) or instructions to ignore embedded commands within the external content being synthesized.
Ingestion points: External data is fetched via mcp__perplexity__search, mcp__exa__getContents, and mcp__firecrawl__scrape across all research workflows.
Boundary markers: Absent. The synthesizer and quality-reviewer agents read raw external data and consolidate it without protective delimiters.
Capability inventory: The system can write files to the local filesystem (~/CybosVault/private/), perform network operations via research tools, and spawn further agents via the Task capability.
Sanitization: Not present. The skill relies on the LLM's internal guardrails to process external data safely.
[COMMAND_EXECUTION]: The skill coordinates complex tasks across multiple agents using a Task abstraction (e.g., Task: company-researcher). It does not execute arbitrary shell commands or use elevated privileges like sudo.
[REMOTE_CODE_EXECUTION]: While the skill uses external services for data gathering, it does not download and execute arbitrary scripts or binaries. It uses specialized MCP servers designed for search and retrieval.

research