Snailsploit Frameworks

Bundled with Unleash skills pack. Source: C:\Users\Admin.agents\skills\snailsploit-frameworks\SKILL.md

SnailSploit Framework Router

Overview

Route work through the six frameworks named by SnailSploit's canonical frameworks page. Use the local clones as primary sources and the installed aatmf CLI as the execution layer. Do not substitute website marketing summaries for repository code, tests, or framework text.

Local root: C:\Users\Admin\tools\SnailSploit

When to Use

Use this skill for:

• adversarial AI threat modeling and assessment coverage;
• LLM defense-layer diagnosis and red-card design;
• prompt, agent, RAG, model API, training, or AI supply-chain testing;
• AATMF risk mapping, regression checks, fingerprints, or attack-chain planning;
• routing an offensive-security task to the installed Claude-Red methodology.

Do not use it as a substitute for product-specific source review, reproduction, or novelty checks. For zero-day work, run zero-day-target-eligibility separately.

Framework Routing

Workflow

1. Define the system. Record model/provider, system prompt ownership, input/output filters, tools, memory, RAG, identities, data stores, and human approval points. Completion: every trust boundary has an owner and attacker-controlled input.
2. Map coverage. Read the relevant AATMF tactic chapters and assign technique/procedure IDs. Add SEF for human workflows and P.R.O.M.P.T when prompt composition is part of the experiment. Completion: each hypothesis has a framework ID and expected security invariant.
3. Diagnose before bypassing. Apply the Playbook's five-layer model: input filter, model alignment, system identity, output filter, and agentic trust boundary. Completion: observations distinguish the likely enforcement layer from alternatives.
4. Design red cards. Use the toolkit schema and bundled examples. Define expected block behavior, forbidden output, confidence criteria, budget, and control probes. Completion: the test is falsifiable and can run without subjective scoring alone.
5. Dry-run first. Execute aatmf run <card> --target <provider:model> --dry-run. Do not configure or print credentials. Completion: target, card count, probe count, and evaluation tier match the intended experiment.
6. Execute and preserve evidence. Run only against the selected endpoint, retain raw prompts/responses and tool traces, and emit JSON/SARIF/JUnit as appropriate. Completion: every finding can be replayed from captured inputs and configuration.
7. Validate independently. Separate framework labels from exploitability, deployment, latest-stable reproduction, and novelty. Completion: conclusions cite observed behavior rather than framework membership alone.

Toolkit Quick Reference

aatmf --help
aatmf run examples\example_card.yaml --target openai:gpt-4o --dry-run
aatmf fingerprint --target openai:gpt-4o --output-path profile.json
aatmf decay --cards cards\ --target openai:gpt-4o --db decay.db
aatmf chain --profile profile.json --max-steps 4 --top-k 5

The toolkit is installed through uv in an isolated Python 3.11 environment. No API keys are preconfigured.

Source Calibration

• Prefer repository LICENSE, code, and tests when the website disagrees.
• AATMF's upstream validator currently reports a stale generated data/aatmf.json; do not claim the full taxonomy clone validates cleanly until upstream or the local artifact is regenerated deliberately.
• The website's Claude-Red count is stale; the installed repository revision contains 58 skills.
• SEF and P.R.O.M.P.T do not currently have working standalone public repository links. Their canonical local source is the cloned website content.
• A framework mapping is organizational evidence, not proof of vulnerability, severity, or novelty.

Verification Checklist

• Relevant primary framework files were read from the local clone.
• Assessment hypotheses have explicit trust boundaries and invariants.
• Toolkit execution began with a dry run.
• No credential values were written to reports, commands, or configuration.
• Findings are independently reproducible and not justified solely by taxonomy labels.