The Agent Skills Directory

[METADATA_POISONING]: The skill package includes a file titled eval_report_methods-reverse-engineer_result.json which contains a self-authored security audit claiming a 'PASS' status. This is a deceptive metadata practice intended to influence the outcome of automated security assessments.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and reconstruct information from external, untrusted sources such as biomedical papers and methods sections.
Ingestion points: Processes PDF text, screenshots of figures, and DOI identifiers as specified in SKILL.md.
Boundary markers: Implements a rigorous labeling system to distinguish between 'explicitly reported', 'inferable', and 'missing' workflow steps.
Capability inventory: Functions are strictly analytical; the skill lacks tools for network exfiltration, file system modification, or command execution.
Sanitization: The instructions do not define specific sanitization or escaping protocols for the content extracted from external documents.
[INDIRECT_PROMPT_INJECTION]: The included evaluation report contains self-referential content asserting a 'safe' verdict, which is a pattern designed to target the security analysis process itself.
[EXTERNAL_DOWNLOADS]: The skill documentation includes an informational reference to a source repository on GitHub (github.com/aipoch/medical-research-skills). This reference is treated as safe as it belongs to the skill's author and uses a well-known service.

methods-reverse-engineer