The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and analyzing untrusted external data, which introduces a vulnerability surface for indirect prompt injection. 1. Ingestion points: The skill instructions direct the agent to read repository instructions, code diffs, pull request descriptions, and linked tickets or acceptance criteria (SKILL.md). 2. Boundary markers: The skill does not provide instructions to use delimiters or explicit 'ignore' commands to safeguard against instructions embedded within the external data. 3. Capability inventory: The agent is tasked with reporting findings and is permitted to modify or post code at the user's request (SKILL.md). 4. Sanitization: No mechanisms for sanitizing, validating, or filtering external content are mentioned in the skill definition.

timotom