In short
AutoCVE is an open-source, AI-driven platform for vulnerability work. Its public project describes multi-agent source code auditing that finds and verifies flaws and drafts CVE reports. Odin Research tracks AutoCVE servers that turn new advisories and patch diffs into HTTP checks, which shortens the time between a fix and scanning for unpatched systems.
Key takeaways
- What it is
- An AI agent platform for vulnerability discovery and verification, published on GitHub under the AGPL-3.0 license.
- How it works
- Coordinated agents handle recon, scanning, triage, finding and verification steps. A web front end talks to a Python API backend.
- Why it matters
- Tools like this cut the time from a published fix to a working check. Attackers and defenders both gain speed.
- What Odin Research saw
- 118 AutoCVE endpoints, 41.4% of all AI vulnerability agents in one study. Seven controllers on dedicated cloud hosting were confirmed.
- What to do first
- Patch faster than checks can be written, and alert on internal hosts that send advisories or code to unknown /api/v1/ services.
What is AutoCVE?
AutoCVE is an open-source platform that uses AI agents to find and verify software vulnerabilities. Its public project calls itself "One-Click CVE Discovery" and covers project selection, code import, agent-driven auditing and CVE report drafting. Odin Research tracks AutoCVE servers as part of a wider class of specialized AI vulnerability agents.
The public project is published on GitHub by the user larlarua under the AGPL-3.0 license. Its README says it is for authorized security research, code auditing and learning, and prohibits unauthorized scanning or testing. It credits DeepAudit for its early architecture.
The project's own claim is that a one-week test found and submitted 30 vulnerabilities across 14 open-source projects, with a top CVSS score of 9.9. Those are self-reported results.
AutoCVE is one of several tools that apply AI agents to the slow, manual parts of vulnerability research. That makes it dual-use. It helps maintainers and researchers, and it can help attackers move faster.
How AutoCVE works
According to its README, AutoCVE runs as a Docker Compose stack:
- a React 18 front end on port 3000;
- a Python FastAPI backend API on port 8000, with Swagger documentation at
/docs; - a PostgreSQL 15 database, with an Adminer management page on port 8080.
An orchestrator coordinates several agents. Recon, scan, triage, finding and verification agents each handle one step. A merge step combines their results. The README lists three audit modes: an enhanced scan, an intelligent audit and a combined comprehensive audit.
Odin Research describes the AutoCVE family it observed in Monograph 20 a little differently. It watches structured feeds, such as the NVD API and vendor code commits. When a fix appears, it turns the patch diff into a proof-of-concept HTTP request and a template for Nuclei, a template-based vulnerability scanner. The report does not tie the servers it observed to one specific repository.
Either way, the output is the same kind of artifact: a repeatable check that tells whether a target is still unpatched. That is a proof-of-concept step, not a finished attack.
Why AutoCVE matters to defenders
Turning an advisory into a working check used to take a skilled person hours or days. Automation shortens that gap. When a patch is published, attackers can test the internet for unpatched systems sooner.
This changes how you should prioritize. Severity scores alone are not enough. Watch for flaws in the KEV catalog and for signs that exploitation is likely, such as EPSS scores. Treat the days after a fix as the riskiest window.
There is also an inside risk. Teams may run tools like AutoCVE themselves. If an internal host sends company code or advisories to an outside AI audit service, that is a data leak, even when the intent is good.
What Odin Research observed
Monograph 20 mapped 285 endpoints running seven families of specialized AI vulnerability agents. AutoCVE was the largest.
- 118 endpoints, 41.4% of the total. They sat on 116 hosts.
- Port 3000 dominates, port 8000 confirms. 105 endpoints answered on port 3000 and 9 on port 8000. All nine confirmed AutoCVE backends listened on port 8000.
- Seven block-ready controllers. Seven confirmed controllers sat on dedicated cloud hosting at Shenzhen Tencent, Hangzhou Alibaba and HostPapa, in China, the United States and Singapore. Two more sat on China Telecom and China Mobile lines, so they go to a short watchlist, not a permanent block.
- A CDN collision. Of 109 candidates, 85 sat behind one shared proxy provider on port 3000 with a
CN=Cloudflarecertificate. A generic web page behind a shared edge says little, and blocking those addresses would break unrelated sites. - Co-hosting. One withheld host paired an AutoCVE console with an AutoHunter parameter miner.
The two ports in Odin's data, 3000 and 8000, match the front end and API ports in the public project's README.
How to detect AutoCVE
- API routes. Odin Research saw AutoCVE accept advisories on a
/api/v1/cve/route. Outbound POST requests withcve_idandpatch_difffields from an internal host are worth an alert. - Task status responses. Responses from ports 8000 or 3000 that carry
autocve_taskandtemplate_statusmarkers point to an AutoCVE backend. - Worker processes. On Linux container hosts, look for
python3,uvicornornucleiprocesses whose command lines nameautocveor a CVE feed option. - Not the port alone. Port 3000 runs countless web apps. Never block on it without a second signal.
Monograph 20 carries Suricata signatures for AutoCVE API calls, a Sentinel KQL hunt, a Sigma rule for unapproved AutoCVE runners and a YARA rule for AutoCVE manifests. They are in the Defense Kit.
How to defend against AI-accelerated vulnerability checks
- Shrink your patch window. Use a vulnerability management process that patches exposed systems within days of a fix.
- Know what is exposed. Map your internet-facing services so that a new advisory triggers an immediate check of your own estate.
- Govern internal AI tools. If your teams run AutoCVE or similar tools, keep them on approved hosts, and block uploads of company code to unknown external services. See shadow AI.
- Block confirmed controllers only. Block the confirmed addresses in Monograph 20 and re-check them after 30 days.
AutoCVE compared with other AI vulnerability agents
| Tool | Role in Monograph 20 | Endpoints |
|---|---|---|
| AutoCVE | Advisory or patch diff to HTTP check | 118 (7 confirmed on cloud) |
| DeepAudit | Source code auditing | 61, all dual-use review |
| Cairn and AutoHunter | Surface discovery: subdomains, routes, parameters | 1 confirmed each on cloud |
| Strix | Sandboxed agent swarms | 5 confirmed on cloud |
| HexStrike AI | General MCP pentest framework | Covered in other monographs |
Frequently asked questions about AutoCVE
What is AutoCVE?
AutoCVE is an open-source platform that uses AI agents to audit code, find and verify vulnerabilities, and draft CVE reports. Odin Research also tracks AutoCVE servers that turn new advisories into HTTP checks.
Is AutoCVE malware?
No. AutoCVE is a dual-use security tool. Its README limits it to authorized research. Defenders track it because attackers can use the same automation to find unpatched systems.
Is AutoCVE legal to use?
Yes, on code and systems you own or are authorized to test. The project is AGPL-3.0 licensed. Its README prohibits unauthorized scanning, penetration testing or security assessment.
Who made AutoCVE?
The public project is published on GitHub by the user larlarua. The README credits DeepAudit for its early architecture and names no company.
What ports does AutoCVE use?
The README lists port 3000 for the front end and 8000 for the backend API. Odin Research found most AutoCVE endpoints on 3000 and all confirmed backends on 8000.
How do I detect AutoCVE?
Watch for requests to /api/v1/cve/ with patch_diff fields, responses with autocve_task markers, and worker processes that name autocve. Odin Research publishes Suricata, Sigma, KQL and YARA rules in its Defense Kit.
AutoCVE vs DeepAudit: what is the difference?
DeepAudit traces untrusted input through source code. AutoCVE focuses on finding and verifying CVE-grade flaws and, in Odin Research's data, on turning advisories into checks. AutoCVE credits DeepAudit for its early design.
Sources
- larlarua, AutoCVE README (English), GitHub. raw.githubusercontent.com/larlarua/AutoCVE/master/README_EN.md (checked 21 November 2026)
- larlarua, AutoCVE README (Chinese), GitHub. raw.githubusercontent.com/larlarua/AutoCVE/master/README.md (checked 21 November 2026)
- ProjectDiscovery, Nuclei overview. docs.projectdiscovery.io/tools/nuclei/overview (checked 21 November 2026)
Related tools and terms
What is CyberStrikeAI?
CyberStrikeAI is a public AI-driven offensive security platform. It spreads its operator interface, API and agent listener across several ports on one host and exposes tools through MCP.
AI offensive toolWhat is PentAGI?
PentAGI is an open-source autonomous penetration testing system that coordinates AI agents and security tools in containers.
AI offensive toolWhat is HexStrike AI?
HexStrikeAI is a public project that exposes many security tools to AI agents through the Model Context Protocol (MCP).
AI security toolWhat is DeepAudit?
DeepAudit is an open-source AI code auditing tool. Application security teams use it to find flaws in source code.
AI security toolWhat is Cairn?
Cairn is a reconnaissance tool that keeps an asset graph built from DNS zones, TLS certificates and HTTP banners.
AI security toolWhat is AutoHunter?
AutoHunter automates headless bug-bounty reconnaissance of subdomains, API routes and query parameters.
Odin Research
Get the rules, not just the definition
Every Odin Research report ships with verified indicators and detection rules you can load today.