Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore

Odin Research / What is / AI tools and agents

What Is DeepAudit? How the AI Code Auditing Tool Works

An open-source team of AI agents that reads source code and hunts for exploitable bugs: what it does, why both defenders and attackers run it, and how to spot it.

By
Odin by Brandefense
Updated
21 Nov 2026
Reading time
7 min read
Type
AI security tool

In short

DeepAudit is an open-source code security auditing platform that uses several AI agents to find vulnerabilities in source code. It plans an audit, maps entry points, traces risky data flows and can test findings with proof-of-concept scripts in a Docker sandbox. AppSec teams use it, and attackers can point it at leaked code.

Key takeaways

What it is
An AGPL-3.0 project on GitHub that combines large language models, static analysis and a Docker sandbox to audit code for security flaws.
How it works
Four agents split the work: an orchestrator plans, a recon agent maps the code, an analysis agent hunts flaws, and a verification agent writes and runs test scripts.
Why it matters
The same tool that speeds up a pre-release review can mine a leaked repository or an exposed .git folder for exploitable bugs.
What Odin Research saw
61 DeepAudit endpoints on 49 hosts in Monograph 20. None was blocked, because none shared a host with a confirmed offensive controller.
What to do first
Remove exposed .git folders and source archives, and alert when internal hosts upload company code to an outside audit service.

What is DeepAudit?

DeepAudit is an open-source platform that uses a team of AI agents to audit source code for security vulnerabilities. It reads a repository, maps its entry points, traces untrusted input to dangerous functions and can test each finding in a sandbox. Security teams use it for code review. Attackers can use it on stolen code.

The project is published on GitHub as lintsinghua/DeepAudit under the AGPL-3.0 license. Its English README calls it a code security auditing platform built on a multi-agent architecture. It accepts code from GitHub, GitLab and Gitea, or as a ZIP upload, and exports reports as PDF, Markdown or JSON.

DeepAudit belongs to a new class of AI agents that specialize in one security job. It does not scan live websites. It reads code and reports likely flaws, such as SQL injection, server-side request forgery and authentication bypass paths.

How DeepAudit works

According to the README, DeepAudit runs four agents:

  1. Orchestrator. Plans the audit strategy and writes the final report.
  2. Recon. Identifies frameworks, libraries, APIs and entry points in the code.
  3. Analysis. Looks for vulnerabilities using retrieval-augmented generation (RAG) and abstract syntax tree (AST) analysis. An AST is a structured map of the code.
  4. Verification. Writes proof-of-concept scripts and runs them in a Docker sandbox, retrying when a test fails.

The backend is Python with FastAPI and PostgreSQL. The frontend is React and TypeScript, served on port 3000 in local setups. Libraries named in the README include LangChain, LangGraph, ChromaDB, LiteLLM and Tree-sitter.

DeepAudit works with many model providers: OpenAI, Claude, Gemini and DeepSeek, Chinese providers such as Qwen, GLM and Kimi, and local models through Ollama. The README warns that code is sent to whichever model provider the user selects. That is a data protection issue for any company that tries the tool on internal code.

The project claims its work has produced 49 CVE IDs and 6 GitHub security advisories across 17 open-source projects. These are the project's own claims, not independently checked here.

Why DeepAudit matters to defenders

Finding a bug in a large codebase used to take days of manual reading. DeepAudit and similar tools shrink that work, for defenders and attackers alike.

  • Leaked code becomes a target list. An attacker who obtains an exposed .git folder, a source archive or a private repository can feed it to an AI auditor and get candidate flaws back. See How to respond to leaked databases and source code.
  • Faster N-day work. Pairing a code auditor with tools that turn advisories into checks can cut the time between a patch and the first probe. Odin Research describes this chain in Monograph 20 as a model, not an observed campaign.
  • Shadow AI risk. A developer who runs DeepAudit with a public model provider sends company code outside the organization. Treat it as shadow AI if no one approved it.

Is DeepAudit legal? Yes. It is open-source software for auditing code you are allowed to read. Using it on stolen code is a different matter.

What Odin Research observed

Odin Research tracked DeepAudit in three monographs. All addresses stay withheld because the tool is dual-use.

  • 61 endpoints, none blocked. Monograph 20 found 61 DeepAudit endpoints on 49 hosts. Of these, 39 were confirmed DeepAudit services (37 on cloud hosts and 2 on dynamic telecom lines), and 22 were candidates.
  • Console plus backend. Port 8000 carried 35 endpoints, mostly FastAPI backends, and port 3000 carried 15 web consoles. Twelve hosts (24.5%) exposed both on the same server. Hosts sat in China (23), the United States (18) and Hong Kong (10).
  • No offensive neighbor. None of the 49 hosts also ran a confirmed offensive controller. From outside, an approved code review server looks the same as an attacker mining a leaked repository, so Odin sends DeepAudit hits to policy review.
  • Sometimes beside offense tools. Monograph 11 found DeepAudit on the same servers as HexStrike AI and ARTEX in some cases, on networks such as Alibaba, AROSSCLOUD and HostPapa.
  • Early sighting. Monograph 04 counted one DeepAudit record, on port 3000 in the United States, among 1,874 AI offense framework records.

How to detect DeepAudit

  • Code leaving the building. The most useful alert is an internal host uploading source code or a .git archive to an outside audit service. Monograph 20 observed code submitted to an /api/v1/audit route.
  • Unapproved backends. Alert when a DeepAudit backend starts on a host that is not an approved AppSec runner, for example a uvicorn process with DeepAudit modules.
  • Your own exposure. Scan your attack surface for exposed .git folders and source archives. These are the inputs an attacker would feed an AI auditor.
  • Not by port alone. Ports 3000 and 8000 host countless web apps.

Odin Research publishes rules for this. Monograph 20 carries a Suricata signature for code sent to DeepAudit (a policy alert, not a block), a Sigma rule that flags DeepAudit backends for dual-use review, and a YARA rule for DeepAudit configuration files. Monograph 11 adds a Suricata signature for DeepAudit consoles. All rules are in the Defense Kit.

How to govern DeepAudit and respond to findings

  1. Approve or block, do not ignore. If AppSec teams want AI code review, run it on named hosts with a local or contracted model, and allowlist those hosts in detections. See the AI Agent Governance Policy Template.
  2. Close source leaks. Remove public .git metadata and backup archives, and rotate any secrets found in them.
  3. Patch faster after disclosure. AI auditors and advisory-to-check tools narrow the window between a fix and an attack. See Vulnerability Management 101.
  4. Do not auto-block dual-use consoles. Use a second signal, such as a confirmed controller on the same host, before treating a DeepAudit server as hostile.

DeepAudit compared with other AI vulnerability agents

ToolJobInputOdin treatment
DeepAuditSource code auditingRepositories and archivesDual-use; policy review
AutoCVETurns advisories or patch diffs into checksPublic advisoriesBlock confirmed controllers
CairnAttack surface mappingTargets and domainsHunt; few confirmed
StrixValidates targets from containersLive targetsHunt; confirmed controllers blocked
HexStrike AIGeneral MCP pentest frameworkLive targetsHunting lead

Frequently asked questions about DeepAudit

What is DeepAudit?

DeepAudit is an open-source, multi-agent AI platform that audits source code for security vulnerabilities and can test findings with proof-of-concept scripts in a Docker sandbox.

Is DeepAudit malware?

No. DeepAudit is a legitimate AGPL-3.0 security tool. It is dual-use: the same analysis that helps a developer fix a flaw can help an attacker find one in leaked code.

Is DeepAudit free?

The software is free and open source under AGPL-3.0. Users pay for whatever model provider they connect, or run local models through Ollama.

Which AI models does DeepAudit support?

The README lists OpenAI, Claude, Gemini and DeepSeek, Chinese providers such as Qwen, GLM, Kimi, Wenxin, MiniMax and Doubao, and local models through Ollama.

Does DeepAudit send my code to an AI provider?

Yes, unless you use a local model. The project warns that code goes to the servers of the provider you select. Check that against your data protection rules first.

How do I detect DeepAudit in my network?

Alert on internal hosts that upload source code to outside audit services, and on DeepAudit backends started outside approved AppSec hosts. Odin Research publishes Suricata, Sigma and YARA rules for these in its Defense Kit.

Should I block DeepAudit servers on the internet?

Not on sight. Odin Research found none of 49 DeepAudit hosts also running a confirmed offensive controller, so it routes them to policy review rather than a block list.

DeepAudit vs HexStrike AI: what is the difference?

DeepAudit reads source code to find flaws. HexStrike AI drives network tools such as nmap and sqlmap against live systems through MCP. One audits code, the other attacks running services.

Sources

  1. lintsinghua, DeepAudit README (English), GitHub. raw.githubusercontent.com/lintsinghua/DeepAudit/v3.0.0/README_EN.md (checked 21 November 2026)

Related tools and terms

All Odin Research explanations · Brandefense glossary

Odin Research

Get the rules, not just the definition

Every Odin Research report ships with verified indicators and detection rules you can load today.

Open the Defense Kit