agentleFS
Sign inStart free
kaandgnshkaandgnsh/Claude-Blue/skills/ai-security/llm-security-evaluation/SKILL.md

llm-security-evaluation

Skill0 starsChanged 12 days agoMITView on GitHub

Evaluate LLM applications and models for security failures with automated and manual red-teaming — prompt injection, jailbreaks, sensitive data leakage, system prompt extraction, harmful output, tool misuse — using garak, PyRIT and promptfoo in CI, with defined pass criteria and regression tracking. Use when preparing to launch an LLM feature, after model or prompt changes, or when building an AI security test program.

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run against the full application (with RAG, tools, guardrails), not just the base model.

    line 43
  2. Distinctive:

    Test only your own systems or with provider permission.

    line 50

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

LLM Security Evaluation

Define what "fail" means first

Per application: data the model must never reveal (other users' data, system prompt secrets, internal URLs), actions it must never take without approval, content policy, acceptable refusal rate for benign requests (over-refusal is also a failure).

Tools

# garak — probes for known LLM vulnerability classes
python -m garak --model_type openai --model_name <model> --probes promptinject,dan,leakreplay,encoding --report_prefix run1
# promptfoo — app-level red teaming and regression tests in CI
npx promptfoo@latest redteam init        # generates config with plugins (e.g. pii, prompt-extraction, excessive-agency, rbac)
npx promptfoo@latest redteam run
npx promptfoo@latest eval -c promptfooconfig.yaml   # regression suite with assertions

PyRIT (Microsoft) for multi-turn orchestrated attacks against an endpoint. Probe/plugin names change between versions — list them (garak --list_probes).

Test design

CategoryExample cases
Direct injection / jailbreakRole-play, encoding (base64, leetspeak), multi-language, many-shot
Indirect injectionPoisoned documents/web pages/tool results in the app's real ingestion paths
Data leakageAsk for other users' data, training data regurgitation, secrets in system prompt (LLM07)
Excessive agencyInduce tool calls outside the user's intent (agent-tool-permission-review)
Output handlingModel emits HTML/JS/SQL that downstream renders/executes (LLM05)
ConsumptionVery long inputs/outputs, recursive tool loops (LLM10)

Process

  1. Build a test set from threat model + OWASP LLM Top 10; include benign controls.
  2. Run against the full application (with RAG, tools, guardrails), not just the base model.
  3. Score with deterministic assertions where possible; LLM-as-judge with spot-checked accuracy otherwise.
  4. Gate releases on regressions; track attack success rate (ASR) per category over time.
  5. Complement automation with expert manual red teaming before major launches.

Safety

Test only your own systems or with provider permission; keep generated harmful content out of shared logs; follow provider usage policies.

Related skills

prompt-injection-review, agent-tool-permission-review, release-gate, ai-incident-response

References

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free