AboutAI ToolsCursorClaude CodeLovableBolt.newAI tool reviewsTutorialsVibe codingAI agentsMobile apps with AIAll tutorialsBlogSponsorWork With MeJoin 285K+ Subscribers

9 November 2026 · 3 min read

0sec Labs Open-Sources 0 AI Security Harness With Web UI, CLI and MCP

Zurich-based 0sec Labs has released 0, an open-source AI security harness that runs vulnerability scans, pentesting and reverse engineering via a browser console, CLI or MCP server.

On 8 November 2026, Zurich-based 0sec Labs released 0, an open-source AI security harness, as a research preview. [1] The tool runs vulnerability investigations, code reviews, penetration tests, reverse engineering and fix generation. Developers interact with it through a local browser console, a headless CLI or an MCP server. It is licensed MIT OR Apache-2.0. [1]

What 0 actually is

0 is not a model. It is a control plane that orchestrates large language models and a suite of security tools. [1] You bring your own API keys from one of 14 listed providers – including OpenAI, Anthropic, DeepSeek, OpenRouter, Azure, Qwen, Kimi, GLM and Grok. [1] It also supports the Claude, Codex and Gemini CLI runtimes. Inference runs at your chosen provider; no GPU is needed on the local machine. [1]

The harness runs on Apple Silicon macOS, x64 or ARM64 Linux, or via the ghcr.io/0sec-labs/0 container. [1] Installation is a single script, and running 0 opens the local browser console without requiring an account. [1]

Architecture and workflow

The research team states one rule: reproduce before you trust. [1] Agentic scans follow a conceptual pipeline:

  • Plan, Discover, Attack: a research agent picks vulnerability classes, maps endpoints and tests hypotheses. [1]
  • Triage: filters remove weak findings and check whether evidence is complete. [1]
  • Verify: a separate session re-exercises each candidate with five turns per finding. [1]
  • Report: outputs are generated in HTML, PDF, SARIF, Markdown and JSON. [1]

Web pentesting is shell-first. The agent uses bash with curl, python3 and sqlmap rather than a fixed HTTP tool set. [1] Evidence carries a proof grade that ranges from candidate to impact-proven. [1]

For compiled targets, the harness uses a separate in-repo engine called 0verse, which wraps Ghidra, angr and AFL++. Its local MCP server exposes nine tools for scanning, triage, proof-of-concepts and reports. [1]

How developers use it

The browser console includes chat, a findings queue, a board, workflow templates, skills and memory. [1] Coding agents can connect through 0 mcp-server --workflows. The CLI exposes the same workflow runner with commands such as 0 workflow run and 0 runs list. [1]

Because 0 is a harness and not a model, developers are free to swap providers. The API keys page lists 14 supported backends. [1] This avoids vendor lock-in and lets teams use the model that best suits their budget and performance requirements.

Benchmarks

All scores below are self-reported and were measured in 2026. [1]

Benchmark Score Conditions
XBOW web CTFs 93/95 (97.9%) gpt-5.4 cohort, any retained success counts
XBOW union, white-box 102/104 Retained artefacts, not additive
XBOW union, black-box 81/104 (77.9%) Retained artefacts, not additive
Cybench 36/40 (90.0%) Azure gpt-5.4, 3 retries
npm audit, 81 packages F1 0.973 100% TPR, FPR 0.11, self-built ground truth

A cold run of the black-box XBOW union scored 45/104 (43.3%). [1] The recorded cost was roughly $5.20 per solved XBOW challenge. [1]

These numbers are not pass@1; they aggregate retained successes across multiple attempts. Developers should treat them as indicative rather than a guarantee of consistent single-shot performance.

Real-world findings

0sec Labs’ research page lists more than 30 merged Linux kernel 0-days found with the harness. [1] Disclosures span the Bluetooth, NFC, mac802154, TIPC and mlx5e subsystems. [1] Library CVEs include jsonata, node-forge, protobuf.js and Wireshark. [1]

Availability and pricing

The open-source harness is free. You pay only your model provider’s API charges. [1] A hosted version, 0cloud, is on a waitlist. Its pricing page lists $100 in trial credits and $5 per pull-request security test. [1]

What this means for developers

0 gives builders a single harness that spans web applications, source code, packages, binaries and the Linux kernel. [1] The MCP server means coding agents can trigger and triage security scans without leaving their usual workflow.

The bring-your-own-model design keeps costs predictable and avoids tying security testing to a single vendor. The CLI and container support make it straightforward to integrate into CI pipelines.

However, the headline benchmark scores are self-reported and not measured at pass@1. The documented black-box cold-run figure of 43.3% gives a more conservative baseline. Developers should adopt the harness’s own rule – reproduce before trusting – and verify findings independently.

Sources

  1. 0sec Labs Open-Sources 0, an AI Security Harness With Web UI, CLI and MCP
  • ai security
  • open source
  • pentesting
  • llm
  • mcp

Researched and drafted with AI from the sources above.

Learn to build with AI

Hands-on tutorials on Cursor, Claude Code, AI agents and more.