Research
Perplexity Open-Sources Bumblebee: A Supply-Chain Security Scanner for Developer Machines
AI coding tools install packages autonomously. Bumblebee checks whether any of them are compromised, scanning npm, PyPI, Go modules, MCP configs, and browser extensions without executing a single line of code.
Perplexity just open-sourced Bumblebee, a read-only security scanner for macOS and Linux that checks developer machines for risky packages, compromised browser extensions, and misconfigured AI tool settings.
The timing matters. In 2026, AI coding tools like Claude Code, Cursor, Replit Agent, and GitHub Copilot routinely install and update packages without manual developer review. Each autonomous installation is a potential supply-chain attack vector. Bumblebee was built to catch those risks before they become breaches.
What Bumblebee scans
Bumblebee covers the full surface area of a modern developer workstation:
Package managers:
- npm, pnpm, Yarn, Bun (JavaScript/Node.js)
- PyPI (Python)
- Go modules
- RubyGems
- Composer (PHP)
Developer tool configs:
- MCP (Model Context Protocol) configurations, the protocol that connects AI assistants to external tools and data sources
- Editor/IDE extensions (VS Code, JetBrains, etc.)
- Browser extensions
What it does not scan: Source code files, runtime processes, or network traffic. Bumblebee is strictly read-only. It reads metadata, lockfiles, manifests, and installed package metadata, without executing code, invoking package managers, or reading your source files.
Why “read-only” matters
Most security scanners need elevated permissions and execute code to do their jobs. Bumblebee’s design is deliberately minimal:
- Zero code execution, it never runs any of the packages it scans
- No package manager invocation, it reads lockfiles directly instead of calling
npm listorpip show - No source file reading, it only examines metadata, not your actual code
- Zero non-stdlib dependencies, the entire tool is written in Go with no external libraries, eliminating the ironic risk of a security scanner introducing its own supply-chain vulnerability
- Single static binary, one file, no installation chain
This means you can run Bumblebee on production developer machines without worrying that the scanner itself could trigger or mask a compromise.
The AI tool config angle
The most novel aspect of Bumblebee is its scanning of AI tool configurations, specifically MCP configs. MCP (Model Context Protocol) is the standard that Claude Code, Cursor, and other AI coding tools use to connect to external data sources, APIs, and tools.
A misconfigured MCP setup can expose credentials, grant excessive permissions, or connect to untrusted servers. As AI coding agents become more autonomous, running multi-step tasks, installing packages, and accessing external services, the attack surface of MCP configurations grows.
Bumblebee flags MCP misconfigurations before they become security incidents.
Connection to Perplexity Computer
Bumblebee was originally built as an internal tool for Perplexity’s own developer systems. Connected to Perplexity Computer, Perplexity’s agentic AI product, Bumblebee can trigger deeper scans automatically whenever a new supply-chain risk emerges in the wild.
This represents an emerging pattern: AI systems that monitor their own security posture. Rather than waiting for a human to run a scan after a CVE is published, the AI agent detects the new risk and initiates a scan across all connected developer machines immediately.
As Perplexity CEO Aravind Srinivas noted, getting agentic AI tools embedded in enterprises requires continuous security engineering, and Bumblebee is part of that investment.
How to use it
Bumblebee requires Go 1.25+ and runs on macOS and Linux:
git clone https://github.com/perplexityai/bumblebee.git
cd bumblebee
go build -o bumblebee .
./bumblebee scanThe output lists all detected packages, extensions, and configs with risk assessments. No account, API key, or cloud connection required, everything runs locally.
The project is licensed under Apache 2.0.
Why this matters for the AI ecosystem
Three reasons Bumblebee matters beyond Perplexity’s own use case:
1. AI-generated code needs AI-grade security. When AI agents install packages autonomously, the traditional model of “developer reviews every dependency” breaks down. Automated scanners like Bumblebee fill the gap between AI speed and human review capacity.
2. MCP security is an emerging concern. As MCP becomes the standard protocol for AI tool integration, misconfigured MCP setups will become a common attack vector. Bumblebee is one of the first tools to address this specifically.
3. Open source builds trust. By open-sourcing Bumblebee, Perplexity lets the community audit the scanner itself, contribute new detection rules, and adapt it for their own environments. For enterprise teams evaluating Perplexity Computer, an open-source security tool is a meaningful trust signal.
For teams building AI-optimized web presence, the broader lesson is that security tooling is becoming part of the AI infrastructure stack, not an afterthought. As LLM crawlers and AI agents interact with more of the web, the security surface expands, and tools like Bumblebee become essential infrastructure.
Related reading
- Best AI coding tools 2026
- Code with Claude May 2026: every announcement
- LLM crawler optimization
- The AI SEO Shift: the complete guide
