AutoCVE is a multi-agent platform that automates the full vulnerability research pipeline — project selection, source auditing, verification, and CVE report drafting — for authorized security researchers.
| Tool | larlarua/AutoCVE — multi-agent automated CVE discovery platform combining tool scanning with LLM-driven source auditing |
| Category | AI-assisted vulnerability research and source code analysis |
| Primary Use | Automating authorized source-code audits and drafting CVE-eligible vulnerability reports against open-source projects the researcher has permission to test |
| Safe Use | Intended for authorized security research, coordinated disclosure, and lab-based auditing of open-source code; findings are meant for responsible CVE submission to maintainers |
| Telemetry Note | All activity is logged locally in PostgreSQL; from a defensive standpoint, the tool's output is disclosure reports rather than covert traffic, making its use observable only through the researcher's own environment and model API calls |
AutoCVE, hosted at larlarua/AutoCVE and written in Python, is an ambitious attempt to compress the entire vulnerability research workflow into a single orchestrated system. Where most AI-assisted auditing tools stop at flagging suspicious patterns, this platform chains project selection, repository import, audit task creation, agent-driven discovery, and finally the generation of a submission-ready CVE report. The README reports roughly 1,428 stars, an AGPL-3.0 license, and a stack built on FastAPI, React 18, PostgreSQL 15, and Python 3.11+, which signals a full web application rather than a command-line script.
The architectural core is a multi-agent design. An Orchestrator coordinates specialized agents — Recon, Scan, Triage, Finding, and Verification — each responsible for one phase of the audit. Recon gathers project context, Scan runs traditional tooling, Triage filters false positives, Finding performs deep source-code mining, and Verification attempts dynamic validation of candidate vulnerabilities before results are merged and finalized. This separation of concerns mirrors how a human researcher works, and the README's architecture documentation (docs/ARCHITECTURE.md referenced as ARCHITECTURE_DESIGN.md) reportedly covers the ReAct loop state machine, tool orchestration, permission protection, and an Agent Runtime — a level of internal documentation unusual for a tool of this kind.
Three audit modes map to different research objectives. Enhanced scanning (Scan → Triage) trades depth for speed, quickly triaging tool output to suppress noise. Intelligent auditing (Finding) is the deep mode aimed at high-value vulnerabilities, explicitly framed for CVE and 0day-style research on code the operator is authorized to audit. Comprehensive auditing combines both pipelines for full coverage. This flexibility is practical: a researcher triaging a large monorepo has very different needs from one hunting a single authorization flaw in a small service.
The Finding agent is described as the platform's centerpiece, purpose-built for CVE discovery. It reads project source directly and combines a ReAct reasoning loop with dedicated tool calls, a Nudge correction mechanism that steers the agent back when it drifts, and a structured termination path called FinalizeFinding that forces output into a format meeting CVE submission criteria. The Nudge detail is telling — it acknowledges the real failure mode of LLM agents in auditing, namely endless exploration or premature conclusion, and builds an explicit corrective mechanism rather than hoping prompt engineering alone suffices.
The interaction model keeps humans in the loop. The README describes the full audit process being preserved as session context, so a researcher can interrogate results afterward: ask the agent to supply additional evidence, explain an attack chain, refine reproduction steps, or extend the analysis. A visualization layer surfaces activity logs, an Agent Tree, tool invocations, stage progress, and preliminary reports — valuable not just for trust but for the documentary rigor that coordinated disclosure to maintainers demands. Findings are auto-submitted by the agents into a structured vulnerability management module with deduplication, and a Skills configuration system lets operators extend each agent's capability boundary with custom abilities.
Deployment is deliberately simple. A single docker compose invocation pulls a prebuilt production compose file, and a source deployment via git clone followed by docker compose up -d --build serves local development. The stack exposes the frontend on port 3000, the backend API on 8000 with Swagger documentation at /docs, and Adminer on 8080 for database inspection. The user flow is linear: configure a model, import a project, create an audit task, track the live audit, manage findings, and export a report.
The project's headline claim is its results table: 30 security vulnerabilities across 14 open-source projects, including a maximum CVSS of 9.9, all reportedly discovered during a seven-day test period. The listed findings are dominated by authorization weaknesses — improper access control, missing authorization, authorization bypass — alongside SSRF, SQL Injection, hardcoded keys, link-following issues, and an RCE. Affected projects named in the README include Chartbrew, Lemmy, JeecgBoot, o2oa, xxl-job, BentoML, typebot.io, OpenReplay, SillyTavern, pimcore, froxlor, and Tautulli, with detailed reports published in a companion larlarua/vulnerability-reports repository. Full reports are the appropriate disclosure channel here, and their existence lends the claims some verifiability.
The vulnerability-type distribution is itself analytically interesting. Access-control flaws dominate because they are exactly the class that static analyzers miss and that LLM agents — which can read route definitions, middleware, and handler logic together — are comparatively good at reasoning about. SSRF appearing repeatedly across Lemmy, xxl-job, typebot.io, and SillyTavern fits the same pattern: taint-style reasoning from user-controlled URL parameters to outbound requests requires cross-file semantic understanding, which is the Finding agent's stated strength. A skeptical reader should still note that severity clustering in web application access logic says as much about the target population — popular open-source web apps — as about the tool.
For authorized professionals, the legitimate fit is clear: auditing open-source code you maintain, are contracted to assess, or are researching under responsible-disclosure norms, then submitting findings through proper CVE channels. The tool's own framing — generate the report, then let the human copy and submit it — keeps the disclosure decision with the researcher rather than automating contact with maintainers. That boundary matters, and it is the correct one for a platform this aggressive in scope.
From a defensive perspective, AutoCVE is worth watching for two reasons. First, maintainers of open-source projects should understand that agent-driven auditing at this level of automation lowers the cost of finding the authorization and SSRF classes of bugs in their code, making disciplined access-control review and outbound-request validation more urgent. Second, blue teams evaluating AI-assisted tooling for their own code review programs can study its Orchestrator/agent decomposition and its ReAct-plus-Nudge control loop as a design pattern applicable to internal secure-SDLC pipelines. Because the platform logs everything it does into PostgreSQL and produces disclosure-grade reports rather than covert infrastructure, its footprint is documentary rather than operational.
Caveats before adopting: the primary README is in Chinese with an English variant available, model configuration is a prerequisite the quick-start only gestures at, and AGPL-3.0 imposes sharing obligations on anyone embedding it into a commercial offering. The 30-CVE figure is self-reported, and CVSS scores as high as 9.9 for an authorization bypass warrant independent verification against the linked CVE records. Still, as a demonstration that a multi-agent pipeline can take a researcher from repository import to submission-ready report, AutoCVE is one of the more complete implementations of the idea currently published.
larlarua/AutoCVE.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.