The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a methodology for ingesting and processing external, potentially untrusted data, which creates an indirect prompt injection attack surface.
Ingestion points: The skill instructions require the agent to read and extract data from .tex files, source code files, result logs (e.g., eval_results.json), and external citation metadata (DBLP).
Boundary markers: The methodology lacks instructions for using delimiters or boundary markers to isolate untrusted content from the agent's primary instructions.
Capability inventory: The skill leverages the agent's file system access and potentially network access (for DBLP lookups) to perform audits. Malicious instructions embedded in a paper or code comment could potentially influence the agent's behavior during the audit.
Sanitization: There are no requirements for sanitizing or escaping the content extracted from external files before it is interpolated into the agent's reasoning or reporting context.

paper-verification