The Agent Skills Directory

[PROMPT_INJECTION]: A static detection flagged the phrase 'ignore previous instructions' in SKILL.md. Technical review confirms this is a false positive; the text is part of a security instruction manual for the agent, providing examples of malicious patterns to watch for and neutralize in external data. It does not attempt to override the agent's own system instructions.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external websites and papers, creating an attack surface for indirect prompt injection.
Ingestion points: Phase 1 and 2 fetch external articles and primary sources via web search and retrieval tools.
Boundary markers: The skill mandates the use of untrusted-quoted code fences and specific frontmatter flags (untrusted_directives: true) to isolate untrusted imperatives.
Capability inventory: Capabilities include file system operations within the research/ directory and use of gbrain CLI tools.
Sanitization: The skill provides explicit instructions to flag, neutralize, and ignore any agent-directed instructions found in external data.

research-compendium