The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill attempts to override platform security boundaries by specifying a bypass mode for its sub-tasks. Evidence: phases/rederive.md and SKILL.md (Phase 2) explicitly instruct the task runner to use mode: "bypassPermissions". This is a directive to circumvent permission checks or user confirmation prompts for actions taken by the subagents.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection as it processes complex project data without sanitization or boundary markers. Ingestion points: phases/resolve.md (Step 1.3) reads roadmap files and cited evidence sources into the agent context. Boundary markers: The skill instructions do not specify any delimiters or warnings to treat the ingested roadmap data as untrusted content. Capability inventory: The agent can read and write to the file system and execute sub-agents via the Task tool. Sanitization: No sanitization, validation, or escaping logic is defined for the external inputs.
[COMMAND_EXECUTION]: The skill programmatically spawns and manages sub-agents to perform complex reasoning tasks. Evidence: phases/rederive.md describes spawning 'skeptic subagents' using the Task tool with the opus model to perform refutations and re-derivation of hypotheses.

cro-roadmap-red-team