0sec Labs’ Open Security Harness Turns AI Into a Pentesting Workbench

Zurich-based 0sec Labs has released 0, an open-source AI security harness designed to bring vulnerability research, code review, pentesting, reverse engineering and fix generation into one workspace. Backed by Y Combinator, the company describes 0 as a software harness rather than a new AI model.

That distinction matters. Users bring their own large language model provider, then 0 coordinates security tools and scoped workflows against authorized targets. Developers can control it from a local browser console, a headless command-line interface or an MCP server connected to coding agents.

The project ships as a research preview under the MIT OR Apache-2.0 license. It runs on Apple Silicon macOS, x64 or ARM64 Linux, and the ghcr.io/0sec-labs/0 container. No GPU is needed because model inference runs at the user’s provider.

A control plane for security workflows

0sec Labs calls 0 a control plane around LLMs and security tools. The research team also describes it as a self-evolving, multi-model harness, with workflows that break an investigation into planning, discovery and attack.

During an agentic scan, a research agent chooses vulnerability classes, maps endpoints and tests possible explanations. Triage filters remove weak findings and check whether the evidence is complete. The verification stage re-exercises each candidate with five turns per finding, helping separate a promising lead from a demonstrated issue.

The architecture documents state one rule: “reproduce before trusting.” That approach carries through to the evidence system, where findings receive a proof grade that ranges from candidate to impact-proven. Reports can be exported as HTML, PDF, SARIF, Markdown or JSON.

Web pentesting is shell-first rather than tied to a fixed HTTP tool set. The agent uses bash with curl, python3 and sqlmap, giving the workflow access to familiar command-line tools while it investigates a target.

Compiled targets use 0verse, a separate engine in the same repository. It wraps Ghidra, angr and AFL++, while the local MCP server exposes nine tools for scanning, triage, proof-of-vulnerability work and report generation.

Browser, CLI and MCP access

Installation uses one script. Running 0 opens the local browser console without an account, and that console includes chat, a findings queue and board, workflow templates, skills and memory.

For coding agents, the connection command is 0 mcp-server –workflows. The CLI exposes the same workflow runner through commands such as 0 workflow run and 0 runs list, so users can move between an interactive console, automation and agent-based development without changing the underlying workflow.

Model access follows a bring-your-own-provider setup. The API keys page lists OpenAI, Anthropic, DeepSeek, OpenRouter, Azure, Qwen, Kimi, GLM and Grok, and the project also supports Claude, Codex and Gemini CLI runtimes. The harness itself is free, but users pay their model provider.

Promising scores with important caveats

0sec Labs presents strong self-reported results from security benchmarks, along with a public record of upstream Linux kernel fixes. The best listed result is 93 out of 95 XBOW web CTFs, or 97.9%, from a gpt-5.4 cohort using any-success grouping.

The figures change when the test conditions become stricter. The XBOW union result reaches 102 out of 104 for white-box testing, while the black-box union reaches 81 out of 104, or 77.9%. A 2026 cold run scored 45 out of 104, or 43.3%.

The project lists 36 out of 40, or 90.0%, on Cybench with Azure gpt-5.4 and three retries. On an npm audit involving 81 packages, it lists an F1 score of 0.973, a 100% true-positive rate and a false-positive rate of 0.11. The recorded cost was about $5.20 per solved XBOW challenge.

Every score in this group is self-reported. The different XBOW totals also reflect different testing conditions, and the retained artifacts in the union results are not additive. Those details make the benchmark figures useful as a snapshot, but they do not turn them into a universal measure of security performance.

From research preview to hosted service

The research page lists more than 30 merged Linux kernel 0-days, with disclosures spanning Bluetooth, NFC, mac802154, TIPC and mlx5e subsystems. It also lists library CVEs involving jsonata, node-forge, protobuf.js and Wireshark.

For now, the open-source harness remains the main offering. A hosted version called 0cloud is on a waitlist, while its pricing page lists $100 in trial credits and $5 per pull-request security test.

That split gives developers a choice between running the control plane themselves and waiting for a managed service. Either way, 0’s central idea is the same: connect an LLM provider to repeatable security workflows, demand evidence for findings and keep the work accessible through the browser, terminal or MCP.

Stay connected via Google News

Follow us for the latest AI News