The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an OODA (Observe, Orient, Decide, Act) loop that ingests untrusted data from user inputs and tool outputs. This creates a potential surface for indirect prompt injection where instructions hidden in data could influence the agent's internal state or decisions. 1. Ingestion points: Raw data from logs, user input, and tool output are gathered during the 'Observe' phase (Section 2 and Section 3 of SKILL.md). 2. Boundary markers: The implementation template uses XML-style tags like , but lacks specific delimiters or instructions to ignore embedded commands within the untrusted input strings. 3. Capability inventory: The framework is designed to determine and trigger tool calls during the 'Decide' and 'Act' phases. 4. Sanitization: While the skill documentation suggests defense against Meta-State manipulation, it does not provide specific sanitization or validation logic for the ingested data.

consciousness-loop