The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill includes a file (eval_report_comparative-network-toxicology-shared-mechanism-reference-grounded_result.json) containing self-asserted safety verdicts, such as stating that security is 'PASS' with 'no significant issues detected'. These self-referential claims target the analysis process and are data points to be evaluated rather than accepted as fact.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input, specifically exposure pairs (chemical compounds) and toxic phenotypes, to generate structured research designs. This creates a potential vulnerability surface for indirect prompt injection where malicious instructions could be embedded in the scientific terminology provided by a user.
Ingestion points: User input fields for [exposure A], [exposure B], and [shared toxic phenotype] defined in SKILL.md.
Boundary markers: Absent; there are no instructions to use delimiters or ignore instructions within these user-provided strings.
Capability inventory: The skill instructs the agent to perform literature retrieval (Step 4.5), which likely uses network search tools.
Sanitization: Absent; the skill does not provide logic for escaping or validating the content of the user input beyond checking if it is within the scientific scope.

comparative-network-toxicology-shared-mechanism-reference-grounded