The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs strong directive language intended to override the agent's default decision-making process. Specifically, it uses phrases like 'ABSOLUTELY MUST', 'not negotiable', and 'cannot rationalize your way out of it' to ensure the agent follows the framework's skill-invocation logic.
[PROMPT_INJECTION]: Establishes an instruction hierarchy where the skill's rules and user instructions are explicitly prioritized over the agent's default behavior, which is a common pattern for defining specific agent personas but technically constitutes a behavioral override.
[PROMPT_INJECTION]: The skill defines an attack surface for indirect prompt injection by prioritizing external content (such as CLAUDE.md and scope.json) as the highest instruction priority without specifying sanitization or boundary markers.
Ingestion points: CLAUDE.md, scope.json, and direct user requests are used to determine engagement tasks and boundaries.
Boundary markers: Lacks explicit delimiters or 'ignore embedded instructions' warnings for the data ingested from the user or external files.
Capability inventory: The skill manages and routes to 31 other technique-focused skills (e.g., recon, exploit-dev, cloud) using the Skill tool.
Sanitization: No explicit sanitization or validation of the contents of CLAUDE.md or scope.json is described.

using-offensive-claude