BugTraceAI is an open-source, self-hosted pentesting suite combining agentic web scanning, an API testing engine, an analyst dashboard, and published security model checkpoints for authorized testing.
| Tool | BugTraceAI/BugTraceAI — modular, self-hosted agentic web pentesting suite with CLI, web dashboard, and API engines |
| Category | AI-driven security automation / web and API penetration testing framework |
| Primary Use | Running authorized agentic web assessments via BugTraceAI-CLI, investigating HTTP/API traffic in BugTraceAI-WEB, and testing API contracts with BugTraceAI-API |
| Safe Use | Educational and authorized security testing only, against applications with explicit written authorization, e.g. the bundled BugStore practice target or owned lab environments |
| Telemetry Note | Self-hosted by design: scan reports and local services stay on the operator's infrastructure, but analysis traffic goes to whatever LLM provider is configured (OpenRouter, Anthropic, Z.ai, Bedrock), so defenders monitoring egress should watch for those API calls and scanning activity from the host |
BugTraceAI positions itself as an open-source, self-hosted web pentesting suite that stitches together agentic scanning, an analyst workspace, an independent API-testing engine, published model checkpoints, and extensible tooling under the Model Context Protocol (MCP). The project is written primarily in Shell at the orchestrator level, carries an Apache-2.0 license, and sits at 342 stars on GitHub at the time of writing. What distinguishes it from the usual wrapper-around-an-LLM scanner is the architecture: rather than one monolithic agent, the suite is split into four independently deployable products — BugTraceAI-CLI, BugTraceAI-WEB, BugTraceAI-API, and BugTraceAI-Launcher — that can run alone or in combination, with a deliberately vulnerable practice target (BugStore) for safe rehearsal.
The README is unusually explicit about the division of labor. BugTraceAI-CLI is the autonomous scanner: a multi-agent pipeline with specialist tools and browser validation, exposed through a Textual terminal workspace, a REST API, and MCP for automation. BugTraceAI-WEB is the analyst-facing dashboard with more than twenty AI security tools, real-time scan monitoring, and a CLI control center. BugTraceAI-API handles contract-aware API discovery and evidence-first API testing as a standalone service, also over REST and MCP. BugTraceAI-Launcher glues it all together with guided deployment, local or Docker runtime profiles, an optional global btai command, and an AI Setup & Repair Assistant. This modularity means a team can adopt the API engine without ever installing the web stack, which is a sensible design for anyone who has watched all-in-one security platforms collapse under their own weight.
The core philosophy stated in the README deserves attention: BugTraceAI combines AI-guided investigation with deterministic security tools, where the AI prioritizes and reasons about hypotheses while tools and evidence validate what is real. That framing matters operationally. Agentic scanners live or die by false-positive discipline, and the pipeline here — visible in the CLI's terminal views labeled Pipeline, Findings, Agents, Timeline, and Logs — is built around evidence collection. The README repeatedly instructs users to inspect the evidence behind each finding and to verify findings manually, and the project's own disclaimer acknowledges that AI output may contain inaccuracies, false positives, or false negatives. For a professional, that candor is a feature, not a hedge.
The scanning pipeline shown in the CLI workspace moves through Recon, Discovery, Strategy, Exploit, Validate, and Report phases, with target configuration, crawl limits, provider selection (F7), and authentication settings (F8) accessible from the TUI itself. The architecture diagram in the README shows the terminal workspace, REST API, and MCP clients all feeding the same CLI scan engine, which in turn drives specialists and browser validation before producing scan reports. The WEB dashboard attaches to either the CLI engine for web scans or BugTraceAI-API for API-target scans, so the browser-based consoles, specialist activity graphs, and report explorers are front-ends over the same evidence store rather than parallel implementations.
The model strategy is where BugTraceAI diverges most sharply from commercial agentic scanners. The project publishes its own security-focused checkpoints on Hugging Face under several families: CORE-Fast for rapid triage and scanner-output review, CORE-Pro for extended investigation and reporting, CORE-Ultra-27B in Q4 and Q6 quantizations for structured PoC generation and tool-calling workflows, and Apex-G4-26B for offensive security reasoning. Crucially, the README clarifies that publishing a checkpoint does not make it the default in every component; each module documents its own supported provider connections. Provider support varies by workflow — the CLI supports OpenRouter, Anthropic, Z.ai, beta Amazon Bedrock, and custom local OpenAI-compatible endpoints, while BugTraceAI-API adds local Ollama and makes AI enrichment optional entirely.
That provider flexibility ties into the project's stated core principles: Privacy-First self-hosted scanning and report storage, with analysis routed to whatever LLM provider the operator configures. Scan reports and local services stay on your infrastructure — a meaningful property for consultancies handling client data under NDAs, since nothing forces findings through a vendor's cloud. The counterpoint is that every AI call still egresses to the configured provider unless you point it at a local Ollama or OpenAI-compatible endpoint, so defenders and red-team leads alike should treat the provider configuration as part of the data-handling boundary for an engagement.
The MCP ecosystem extends the platform beyond its built-in capabilities. The project ships a reconftw-mcp integration exposing ReconFTW to the agent framework, with Kali Linux agents mentioned as integrated as well. This is the standardization bet: rather than hard-coding tool integrations, BugTraceAI treats external security tools as MCP-addressable services that agents can route work to. For operators already invested in ReconFTW workflows, that means existing recon assets can be absorbed into the agentic pipeline instead of being replaced by it.
Practice and validation get first-class treatment. BugStore is a deliberately vulnerable, full-featured shop application seeded with 32 planted OWASP vulnerabilities, used throughout the project's own demos and walkthroughs. The README is careful to note that report counts visible in the web dashboard screenshots describe example sessions against BugStore, which tells you the marketing artifacts are reproducible in a lab rather than cherry-picked from live targets. For anyone evaluating the suite, BugStore is the correct starting point: it gives a bounded, authorized target where the agents' findings can be checked against a known vulnerability inventory.
The project's credibility claims are concrete rather than vague. The README lists three CVE disclosures attributed to the platform — CVE-2026-27479 against Wallos (CVSS 7.7), CVE-2026-27470 against ZoneMinder (CVSS 8.8), and CVE-2026-27834 against Piwigo (CVSS 7.2) — plus stage presentations at RootedCON 2026 in Madrid, HKOSCon 2026 in Hong Kong, and DEF CON 34 in Las Vegas, with a public product demo video linked from the README. A sample report is also available through the project's demo site. Disclosed CVEs in real open-source products are a reasonable proxy for whether an automated pipeline actually finds bugs, and it is the kind of evidence most competing tools do not publish.
The README also sketches a CI/CD integration proposal and points to a Launcher-driven quick start, though the truncated text does not enumerate the exact installation commands; the guided workflow centers on choosing WEB, CLI, API, or a combination, then selecting local or Docker runtime profiles. The CLI itself is versioned at 4.0.16-beta in the current screenshots, and the README notes an offline demo mode with sample data built into the terminal workspace — a thoughtful touch that lets you evaluate the TUI without pointing it at anything.
For authorized professionals, the fit is clear: bug-bounty hunters working in-scope programs, internal pentest teams with written authorization, and defensive researchers studying how agentic pipelines structure findings will all find something here. The mandatory framing is honest — educational and authorized testing only, explicit written authorization required, AI output treated as hypotheses rather than verdicts. The suite's own language that it is not a substitute for professional auditing, combined with its evidence-first reporting model, reads as a tool designed to augment analyst judgment rather than replace it, which is exactly the posture a senior operator should want from AI-assisted offensive tooling.
BugTraceAI/BugTraceAI.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.