The Agent Skills Directory

Adversarial Audit Methodology: The skill instructs the agent to 'ignore existing assumptions of safety' and 'treat all inputs and boundaries as untrusted.' This language is an application of a standard adversarial research mindset intended to ensure the agent identifies vulnerabilities in the target codebase, rather than a bypass of the agent's own safety guardrails.
Indirect Prompt Injection Surface: As a code analysis tool, the skill is designed to ingest and process arbitrary source code, which acts as untrusted data. This introduces a potential surface for indirect prompt injection if the audited code contains instructions meant to influence the LLM.
Ingestion points: Source code files are read based on the target_files listed in workspace/plan.json and the metadata in kb_references.
Boundary markers: The skill relies on the agent's own context management, though it does not explicitly define unique delimiters for the ingested code content.
Capability inventory: The skill generates JSON finding files in workspace/findings/, executes grep commands for call-site discovery, and runs a local Python script (workspace/helpers/query_structural_index.py).
Sanitization: No specific sanitization or filtering is described for the raw source code being analyzed.
Local Script and Tool Execution: The skill invokes grep and a local helper script to perform structural analysis and prioritize review targets. These operations are restricted to the local workspace and are routine for automated security research workflows.

mantis-researcher