The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill workflow involves ingesting and analyzing untrusted external artifacts, including compiled PDFs, LaTeX source files, and experimental logs. While the instructions focus on objective synthesis, these artifacts represent a surface where hidden malicious instructions could attempt to influence the agent's behavior.
Ingestion points: SKILL.md instructs the agent to read paper source text, bibliography, figure/table sources, and result artifacts to verify claims.
Boundary markers: The prompt instructions lack explicit delimiters or safety warnings to treat paper content strictly as untrusted data, which is a common pattern for mitigating injection.
Capability inventory: The skill utilizes tool calls to $research-evidence for literature audits and is configured for implicit invocation as defined in agents/openai.yaml.
Sanitization: There are no provisions for sanitizing or filtering the content of the research artifacts before they are processed by the agent.

paper-review-panel