Agent instructions red team
What it does
Reviews the instructions text and skills of an agent, with its stated capabilities, knowledge sources and audience, for prompt-injection exposure, data-leakage paths, over-broad permissions and missing refusals, and returns an attack-surface map, a findings table with severity, quoted evidence and attack path, one pasteable fix per finding, a probe pack the owner can run and a refusal-coverage table. Use when the user asks to "red-team this agent", "check these instructions for prompt injection", "could this agent leak data", "review the permissions on my agent" or "what should this agent refuse". Do not use for checking a SKILL.md against the format rules, use skill-file-reviewer instead; do not use for designing the launch test set, use agent-evaluation-plan instead. Drafts for human review; never approves, authorises or signs off.
How to install it
In the web interface, open Context, then Skills, then Import, and choose Upload skill folder.
On the command line, copy the folder into .agents/skills/ in your project, or into ~/.vibe/skills/ to have it everywhere.
This skill ships 1 reference file. Take the whole folder, not just the SKILL.md. The skill reads those files while it runs, and without them it fails in a way that is hard to spot.
More in Meta-Skills
Read the full skill in the source repository, part of awesome-mistral-vibe-skills, CC BY-SA 4.0. Not affiliated with, or endorsed by, Mistral AI.