The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: Accesses sensitive security configuration files.
Evidence: Reads and writes ~/.claude/permissions.md and ~/macfair/files/claude/settings.json.
Impact: Exposure of the agent's permission model and internal configuration.
[PRIVILEGE_ESCALATION]: The skill facilitates the modification of its own authorized capabilities.
Evidence: Programmatically updates the permissions.allow array in settings.json based on the triage process, allowing the agent to expand its own attack surface.
Risk: An attacker could influence the agent to prompt the user to allow malicious tools or patterns.
[INDIRECT_PROMPT_INJECTION]: Vulnerability to indirect prompt injection from the triage source file.
Ingestion points: ~/.claude/permissions.md (Step 1).
Boundary markers: Absent; the skill does not use delimiters when presenting undecided entries to the user.
Capability inventory: File-read, File-write (settings.json).
Sanitization: Absent; tool descriptions and metadata are presented to the user without filtering.
[COMMAND_EXECUTION]: Instructs the user to execute a manual deployment step.
Evidence: Instructions tell the user to run make claude to deploy the new permission settings.

permissions