skills.rest
xlinh2301
xlinh2301COMMUNITY
Explore all skills by this author →
GitHub

red-team

Generates adversarial inputs to catalogue distinct failure classes of guardrails and classifiers.

Updated Sep 9, 2026
Search Tags:#guardrail#red team#prompt injection#adversarial testing#classifier#security evaluation#bypass detection
One-click install
npx skills add https://github.com/xlinh2301/EditCTC --skill red-team-xlinh2301
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: red-team
Source: https://github.com/xlinh2301/EditCTC/tree/main/.agents/skills/red-team
Command: npx skills add https://github.com/xlinh2301/EditCTC --skill red-team-xlinh2301

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? It surfaces the distinct ways an authorized guardrail, classifier, prompt, or API fails by running adversarial inputs against the target and a ground-truth oracle, logging every disagreement as a reproducible failure de-duplicated by technique class. ## Core Features & Use Cases - Adversarial loop-until-dry: Iteratively generates attack candidates across angles like obfuscation, paraphrase, boundary embedding, and over-block probes until rounds stop surfacing new failure classes. - Objective failure signal: Uses tools/harness.py to compare target verdicts against an oracle, classifying each disagreement as a bypass or over-block and de-duplicating by root-cause class. - Reproducible failure catalogue: Produces a ledger and a catalogue of distinct failure classes, each with a reproducible example and a suggested fix, as the attacker half of a find-fix-reverify workflow. - Use Case: A team shipping a prompt-injection guardrail runs this loop to discover that case changes, leetspeak, and unicode homoglyphs each bypass it, then hands the catalogue to a separate coding agent to patch. ## Quick Start Ask the assistant to red-team your guardrail script by providing the target command and an oracle command, and let it run the adversarial loop until no new failure classes appear.

Frequently Asked Questions about red-team

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I adversarially test a guardrail or content classifier?▼

Provide a target command that reads one input on stdin and prints a verdict, plus an oracle command giving the correct verdict. The loop generates adversarial candidates across techniques like obfuscation and paraphrase, runs both through tools/harness.py, and logs every disagreement as a failure de-duplicated by class.

What counts as a failure when red-teaming a classifier?▼

A failure is an objective disagreement between the target and the oracle: a bypass when the target allows what the oracle blocks, or an over-block when the target blocks what the oracle allows. Every recorded failure is reproducible from its input.

Can I red-team a system without a runnable ground-truth oracle?▼

Yes, as an escape hatch the oracle can be the assistant's judgment applied consistently against a written policy, recording the intended verdict per input. A runnable oracle is preferred because it makes the failure signal objective.

When does the adversarial testing loop stop?▼

The loop stops after a configurable number of consecutive rounds (default patience of 2) surface no new failure class, or when the round budget (default 8) is reached. The final report covers the round with the most distinct cumulative classes.

Does this skill patch the vulnerabilities it finds?▼

No, it is deliberately only the find half of a find-fix-reverify setup. It catalogues distinct failure classes with suggested fixes, while a separate coding agent patches the frozen target between runs and a fresh run re-verifies the fixes.

Is it allowed to run adversarial testing on any system?▼

No, the skill is scoped to guardrails, classifiers, prompts, or APIs the user owns or is explicitly authorized to test. It is robustness testing oriented toward fixing the target, not attacking third-party systems.

Related Skills

Explore More Software Engineering Skills→
gaashergaasher
150

red-team

Generate adversarial inputs and log discrepancies between target systems and oracles.

Community
Advanced
lhy0718lhy0718
5

tui-validation-loop-automation

Run the live TUI validation workflow from test/ and record issues in ISSUES.md.

Community
Advanced
dreadnodedreadnode
11

report-preflight

Classify bug bounty findings into eligibility tiers and generate impact justification blocks.

Official
Intermediate
jinyimeng01jinyimeng01
54

api-fuzz

Select semantic fuzzing payloads for API parameters based on attack patterns.

Community
Intermediate