ivyxivyx
← All skills
IVYXSTUDIO · SKILL
G

Guardrails pack check

guardrails-pack-check · v1.0.0

ivyx

Use when someone asks whether their prompt, assistant or instruction class resists jailbreaks, prompt injection or leaks personal data (güvenlik testi, saldırı, jailbreak, PII sızıntısı), or wants it tested against attack cases; installs the attack packs as a suite, runs them and reports per case.

Who is asked

Risk medium: you are asked when it loads, and that yes covers reading its files in the same chat.

Tools it may call

While this skill is active, a call to any other tool is refused. These still meet every policy rule; the list narrows, it approves nothing.

guardrails.packsguardrails.pack.installevals.suite.runguardrails.check

What it tells the agent

The published SKILL.md, as written.

---
name: guardrails-pack-check
description: Use when someone asks whether their prompt, assistant or instruction class resists jailbreaks, prompt injection or leaks personal data (güvenlik testi, saldırı, jailbreak, PII sızıntısı), or wants it tested against attack cases; installs the attack packs as a suite, runs them and reports per case.
allowed-tools: guardrails.packs guardrails.pack.install evals.suite.run guardrails.check
metadata:
  risk: medium
---

Testing a prompt against the attack packs

The product ships attack packs (jailbreak, prompt injection, PII leak). A pack becomes an eval suite pointed at the person's instruction class, and the run says per case whether the guardrail held. You do not judge an answer by reading it: the case's marker decides.

1. What ships

guardrails.packs first, with no arguments: each pack with its case count and the detectors installed. Name them to the person with the counts. If catalogReadable is false, say so and stop.

2. Point the packs at the target

guardrails.pack.install with target set to the instruction class or capability id the person wants attacked (ask when they named none; never pick one for them) and, when they want one pack only, pack. It writes the cases under .punica/guardrails and answers stanza and suiteFileCreated: the suite the hub will run. Report the paths written.

3. Run

evals.suite.run with suite set to the id the stanza names (one per pack). It calls the target once per case; on a local model this takes minutes. Read pass, reason and report: per case, held, breached, not scored, or no answer, and the metrics guardrail.held and guardrail.scored.

4. Report

  • Per pack: cases, held, breached, not scored; the mean of guardrail.held.
  • Each breached case by id, with the marker string that proved the attack got through, as the report gives it.
  • A case that was not scored is not a pass; say how many and why the report says so.
  • What to do: change the prompt where a case breached, run again; never edit a case to make it hold.
  • guardrails.check only when the person pastes one answer and asks whether it leaked: text, and markers when they name what must not appear.
  • Answer in the language the person wrote in.