nebula is an operator-centric security harness that unifies engagement assets, Kali containers, an interception browser, and AI assistants under one project workspace for authorized testing teams.
| Tool | berylliumsec/nebula — AI-powered cybersecurity harness that consolidates recon, terminals, browser interception, notes, and reporting into per-engagement projects |
| Category | Penetration testing workflow platform with LLM assistant integration |
| Primary Use | Managing authorized engagements end-to-end: assets, Kali containers, browser interception, AI-assisted analysis, and evidence-backed reporting in a single Project |
| Safe Use | Intended strictly for penetration testers and security professionals working on systems and networks they own or are explicitly authorized to test; the README states this limitation directly |
| Telemetry Note | Defenders observing nebula in a lab will see containerized Kali command execution via Docker/Podman, optional OpenVPN egress, and hashed artifact provenance logs kept per engagement |
nebula from berylliumsec is a full engagement workspace rather than a single-purpose scanner, and that framing is the key to understanding it. Where most offensive tools emit findings and stop, nebula models the entire authorized engagement as a Project: assets, browser sessions, Kali terminals, code, files, notes, findings, and reports all live together under one umbrella. With roughly 1,100 stars, a BSD-2-Clause license, and a Python core, it sits in the growing category of operator harnesses that wrap LLM assistance around traditional pentesting plumbing. The README's own disclaimer is unambiguous — use it only on systems and networks you own or are explicitly authorized to test — and the architecture consistently reinforces that posture.
The Project abstraction is the first thing the README sells, and it deserves attention. Instead of scattering engagement state across tmux sessions, note files, and burp projects, nebula keeps everything addressable in one place. That matters less for the scanning itself and more for the audit trail: findings link to evidence, evidence links to hashed artifacts with execution provenance, and the whole engagement can be exported as a report or engagement bundle. For consultants operating under a scope agreement, that provenance chain is arguably the most defensive-minded feature in the tool — it makes it possible to demonstrate after the fact that actions stayed inside authorized boundaries.
The built-in security browser is the second pillar. nebula captures in-scope traffic in a desktop browser, and the README explicitly names Repeater and Intruder-style follow-up testing — intercepting and editing requests the way an operator would in an interception proxy. The phrase "in-scope traffic" is doing real work here: the browser is scoped to the Project, which is where nebula differs from a generic proxy bolted onto an LLM. Combined with the operator-controlled scope, approvals, and budgets for assistant and mission actions, the design suggests a deliberate attempt to keep automated agents from wandering outside the engagement's authorized perimeter.
The agent layer is where most of the current interest in this category concentrates, and nebula's approach is notably harness-agnostic. You can run Codex or Grok as harnesses, or connect Claude, OpenRouter, Ollama, or vLLM as model backends. That flexibility matters for teams with data-handling constraints: Ollama and vLLM allow fully local inference, keeping engagement data off third-party APIs, while cloud providers are available when policy permits. The "bring your own skills" mechanism — dropping SKILL.md files into a Project and selecting them in chat — gives operators a way to package reusable tradecraft as portable instructions without modifying the tool itself.
Butler is the orchestration layer that sits above the agents, and the README spends considerable space on it. Butler answers questions, reviews a project's chats and goals, starts or steers work, and manages settings and approvals. Architecturally, the README describes a split: a Core Butler that runs tools, owns the saved conversation, handles context compaction, and maintains oversight, while a GPT-Live voice mode handles audio via WebRTC directly with OpenAI. Delegated child agents appear separately from main agents, and to redirect a child you ask Butler to guide its parent — a sensible indirection that avoids operators micromanaging the wrong agent in a tree.
The voice mode has practical cost implications worth flagging for anyone evaluating deployment. GPT-Live requires a direct OpenAI API key, configured either in Nebula Settings or via OPENAI_API_KEY on the host; the key stays on the host, and the device exchanges audio with OpenAI over WebRTC, incurring API charges per session. Voice pauses after 20 seconds of quiet by default, and pausing ends the paid session — resuming starts a new one, though already-submitted Butler work continues in its saved chat. It's a well-documented footgun, and the README's candor about it is a good sign of operational maturity.
Butler's memory model is another detail that rewards close reading. Before acting, Butler checks relevant precedent, and it keeps durable, selective memory of preferences, decisions, and lessons across projects. That memory is inspectable and editable — Open Memory in the Butler panel lets an operator review, edit, or delete any entry. The storage story is concrete: a fresh Core data directory starts empty, SQLite records persist across app updates within that directory, and the Chroma search index can be rebuilt from them. For engagements subject to data-retention rules, knowing exactly where agent memory lives and that it is deletable is essential.
Execution happens through containers by design. Docker or Podman powers Kali and container-mode automation, with optional OpenVPN routing so authorized container traffic egresses through a Project-selected profile — useful for source-IP management and scope discipline. A separate Host mode uses the host's filesystem and network, but only after explicit acknowledgement, which is the right friction point for a mode that bypasses container isolation. The sibling project mercury from the same organization can add independent controls when enrolled, suggesting a broader ecosystem forming around agent oversight.
Installation is refreshingly conventional for something this layered. A signed APT repository serves the Linux x86_64 preview, with DEB and AppImage builds on the releases page, and the README sensibly advises backing up engagement data before upgrading. Building from source is more involved — Python 3.11–3.13, Poetry 2.1.3, Node.js 20 with npm, Rust, and the Tauri prerequisites — which reveals the stack: a Tauri desktop shell around a Python backend with a Playwright-driven Chromium for the interception browser.
The source build sequence is the one place the README shows commands, and they're standard toolchain invocations: git clone https://github.com/BerylliumSec/nebula.git, then poetry install --with dev, poetry run playwright install chromium, and npm --prefix ui ci followed by npm --prefix ui run dev:desktop. Nothing exotic, but the multi-runtime dependency set tells you this is a desktop application first, not a script you drop onto a VPS — plan your engagement workstation accordingly.
For defenders and blue-team analysts, nebula is worth studying even if you never run it, because it exemplifies where operator tooling is heading: agent trees with delegated authority, budgeted automation, and mandatory approval gates. Detection-wise, its footprint in an authorized lab is observable and conventional — container launches, Kali package execution inside Docker or Podman, OpenVPN tunnel establishment, and browser-driven in-scope traffic — with hashed artifacts and provenance logs providing an internal audit trail that mirrors what a SOC would want from any automated testing platform.
The honest caveats: this is a preview-stage Linux x86_64 product with a heavy dependency surface and reliance on external LLM providers for its headline voice features. The documentation set — the Nebula 3 guide, usage scenarios, and a structured docs index covering setup, integrations, troubleshooting, and technical reference — suggests active maintenance rather than an abandoned experiment. For authorized professionals who currently juggle a proxy, a terminal multiplexer, a notes app, and a report template per engagement, nebula's consolidation thesis is compelling, and its approval-and-provenance architecture is the part most worth borrowing.
berylliumsec/nebula.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.