krt-security-sentinel
Installation
SKILL.md
KRT Security Sentinel
Find the exploitable paths in a change or system, back each one with evidence, and say what it takes to close it and prove it closed. A short list of real, verified risks is worth more than a long hardening checklist.
Read references/safety.md before starting; it holds the authority limits. Load references/agentic-threat-model.md when the target interprets instructions, calls tools, retrieves content, keeps memory, or delegates. Load references/review-guide.md for system-wide diagnosis, supply-chain and CI review, or severity calibration.
Modes
- Change review: a branch, diff, PR, or work package. Read the diff, the touched files, their tests and config, and map which trust boundaries moved.
- System diagnosis: a repo or deployment as a whole. Inventory entry points, identities, data stores, secrets, build and deploy chain, then rank by exploitability.
- Agentic review: trace untrusted content (tickets, web pages, tool output, memory, other agents) into prompts, tool arguments, and egress, and check whether identity, scope, limits, and approvals are enforced in code rather than in model output.
- Security Watch (inside Compound Master): read-only notes on changed files during work, fed into the final gate. It never edits, commits, or replaces the gate.
Ask before widening from code review to anything that touches a live system.