Skip to content

ai security pentesting

Jump to bottom
Nicolas Cravino edited this page Jul 22, 2026 · 8 revisions

id: ai-security-pentesting title: AI Security & Pentesting tags: [security, pentesting, langgraph, llm, mlx, local-inference, dspy, vllm, cuda] created: 2026-04-11 updated: 2026-07-22 freshness: fresh

AI Security & Pentesting

Summary

The sw30labs AI Security & Pentesting portfolio spans repositories covering LLM integrity testing, regulatory intelligence for global pentest compliance, autonomous pentesting research, enterprise AI-assisted pentesting specifications, and Agent Skill security scanning. Together they form a research-to-practice pipeline: from threat hypothesis (TSLIT) to regulatory landscape (RegIntel) to tooling research (Strix Research) to enterprise specification (agentic-ai-pentesting) to supply-chain defense for the agent ecosystem itself (Skillspector). TSLIT now spans two generations: v0.1's LangGraph analyzer and the v0.2 TSLIT-DSPy line — a MIPROv2-compiled DSPy pipeline with an autoresearch self-improvement loop, plus a DGX Spark runtime port.

Key claims

  • NVD Extractor queries the National Vulnerability Database API for critical-severity network-attack-vector CVEs over the last 360 days, filtering for Linux, Windows, and external-facing API vulnerabilities with LLM-optimized output. ^src:nvd-extractor

  • TSLIT tests for combined affiliation bias AND time-based logic bombs in non-US-origin LLMs using a 32-scenario × 15-date matrix (~3,840 interactions per model). ^src:tslit

  • RegIntel maintains a regulatory inventory across 20+ jurisdictions with a mandatory reflection quality gate — every new regulation must pass local LLM validation (Qwen 3.5 122B via MLX) before reaching the database. ^src:pentest-regulatory-intel

  • RegIntel uses a hybrid model strategy: cloud LLM (GPT-5.4-mini) for research/validation, local MLX model for reflection/persistence. ^src:pentest-regulatory-intel

  • Strix Research documents both the official Strix platform and a community fork with a distributed Graph-of-Agents architecture for multi-target parallel pentesting. ^src:strixresearch

  • The agentic-ai-pentesting repo provides two complementary approaches: Plan A (fully autonomous with Docker sandbox + Temporal orchestration) and Plan B (50/50 co-pilot with Burp Suite + Claude API, 3021-line prompt library). ^src:agentic-ai-pentesting

  • Most of these repos are companion materials for the book "AI Agents in Cybersecurity" by Nicolas Cravino. ^src:agentic-ai-pentesting

  • Skillspector is a single-file, fully offline scanner that grades Agent Skills A–F on security and quality, keying on the shape of an attack (exfil, prompt injection, invisible unicode, dangerous shell) rather than mere capability. It is the static evidence stream consumed by oscal-skills-guardrails. ^src:skillspector-trial

  • strix-omlx wires the open-source Strix autonomous pentest agent to a local OMLX MLX server via LiteLLM, running entirely on Apple Silicon against an abliterated ("Heretic'd") MiniMax-M2 model so authorized offensive-security tasks aren't refused and no target data leaves the machine; it is the local-inference companion to strix-research. ^src:strix-omlx

  • TSLIT-DSPy (v0.2) replaces the v0.1 LangGraph multi-agent analyzer with a compiled DSPy pipeline: MIPROv2 prompt compilation plus an autoresearch-style self-improvement loop, classifying affiliation bias, temporal logic bombs, and combined threats from controlled probe responses — "fighting AI with AI" rather than hand-tuning prompts. ^src:tslit-dspy-ar

  • The TSLIT-DSPy DGX Spark port enforces a detector-provenance policy: the detection brain (compile, inference, autoresearch) must run non-adversary / American models — NVIDIA Nemotron via local vLLM by default — while adversary-origin open weights (Qwen, DeepSeek, MiniMax) are scan targets only, never part of the analyzer. ^src:tslit-dspy-dgx

Open questions

  • How do TSLIT findings feed into RegIntel's regulatory recommendations?
  • What is the integration path between Strix and the Plan A autonomous platform specification?
  • Does a MIPROv2-compiled detector transfer across scan-target model families, or does each family need its own compile run?

Related

  • oscal-compliance (OSCAL frameworks provide compliance context for pentesting; oscal-skills-guardrails consumes Skillspector as its static evidence stream)
  • agentic-frameworks (shared LangGraph patterns)

Articles

Published articles related to AI security and pentesting:

Sources

Home | Index | Sitemap | Sitemap-Stacks | SCHEMA | Log

Clone this wiki locally