The Agent Skills Directory

Adversarial Auditing Context: The skill uses language such as "ignore existing assumptions of safety" and "full freedom and autonomy" within its exploratory investigation logic. While these phrases are often associated with prompt injection, here they are contextually restricted to the code auditing process, directing the agent to maintain a zero-trust posture toward the code being reviewed rather than bypassing its own system safety constraints.
Data Ingestion and Surface Area: The skill processes external data including plan.json, source files, and knowledge base references. This creates an interface for indirect prompt injection where instructions embedded in the audited code could attempt to influence the agent's behavior. However, this is an inherent part of code analysis functionality, and the skill uses structured output (JSON UUID-based files) to manage and isolate findings.
Sub-Agent Delegation: The instructions include directives for spawning waves of sub-agents to parallelize triage and deep-dives. This utilizes multi-agent orchestration features to improve efficiency and does not involve unauthorized code execution or privilege escalation.

mantis-researcher