Updated May 26

Share this article

More on This Story

SpaceX Acquires Cursor for $60 Billion in First Post-IPO Power Move

Jun 17, 2026

SpaceX Acquires Cursor for $60 Billion in First Post-IPO Power Move

SpaceX is buying AI coding startup Cursor for $60 billion in an all-stock deal, its first major acquisition since going public. The deal gives Elon Musk's company access to Cursor's 1 million-plus developers and a foothold in the fast-growing AI coding tools market.

spacexcursoranysphere

Trump Administration Forces Anthropic to Pull Fable 5 and Mythos 5 Offline

Jun 16, 2026

Trump Administration Forces Anthropic to Pull Fable 5 and Mythos 5 Offline

The U.S. Commerce Department ordered Anthropic to block foreign access to its newest Fable 5 and Mythos 5 AI models, forcing a total shutdown. The directive, triggered by an Amazon security paper and a three-word jailbreak prompt, has sparked a global sovereign AI backlash from the UK, France, and Canada.

anthropicfable-5mythos-5

Judge Dismisses xAI Trade Secret Lawsuit Against OpenAI With Prejudice

Jun 16, 2026

Judge Dismisses xAI Trade Secret Lawsuit Against OpenAI With Prejudice

A federal judge dismissed xAI's trade secret lawsuit against OpenAI with prejudice, ruling it would be 'futile' for Elon Musk's company to continue. The dismissal marks Musk's second court defeat against OpenAI in four weeks and ends a case that alleged OpenAI sought Grok 4 trade secrets during recruiting.

xai-lawsuitopenai-lawsuittrade-secrets

Related News

Elastic Agent Memory Demo Shows Why Long Context Is Not Enough

Jun 18, 2026

Elastic Agent Memory Demo Shows Why Long Context Is Not Enough

Elastic published an open-source agent memory demo that turns long-running AI context into a search and retrieval architecture problem.

agent-memoryelasticsearchai-agents

GPT‑NL: a sovereign language model for the Netherlands

Jun 18, 2026

GPT‑NL: a sovereign language model for the Netherlands

GPT‑NL: a sovereign language model for the Netherlands is drawing attention from AI builders and teams evaluating tool adoption, workflow automation, and source-backed decision making.

ai-toolsautomationdeveloper-tools

Launch HN: Adam (YC W25) – Open-Source AI CAD

Jun 18, 2026

Launch HN: Adam (YC W25) – Open-Source AI CAD

Launch HN: Adam (YC W25) – Open-Source AI CAD is drawing attention from AI builders and teams evaluating tool adoption, workflow automation, and source-backed decision making.

ai-toolsautomationdeveloper-tools

Perplexity Open-Sources Bumblebee to Scan Developer Machines for Supply-Chain Threats

Developer Security

Perplexity Open-Sources Bumblebee to Scan Developer Machines for Supply-Chain Threats

Perplexity has open‑sourced Bumblebee, a read‑only security scanner that checks developer machines for compromised packages, browser extensions, and AI tool configurations without ever executing potentially malicious code. The tool, written in Go with zero external dependencies, already protects the systems behind Perplexity Search, Comet browser, and Computer agent.

What Bumblebee Actually Does

When a supply‑chain vulnerability surfaces, security teams face one urgent question: which developer machines are exposed right now? Existing tools miss the answer. Software Bill of Materials scanners cover build artifacts. Endpoint detection products watch running processes. Neither checks the local developer state — lockfiles, package manifests, extension configs, and AI tool settings scattered across a laptop’s filesystem.

Bumblebee fills this gap. It’s a read‑only inventory collector for macOS and Linux developer endpoints, written entirely in Go with zero non‑standard‑library dependencies. On each invocation, it performs a single scan and outputs structured findings as newline‑delimited JSON — no persistent daemon, no process monitoring, no network inspection.

Read‑only by design Reads metadata files directly — lockfiles, manifests, installed package records. Never invokes npm, pnpm, pip, or any package manager. As ² explains, npm packages can carry postinstall scripts that run automatically on install — a scanner that invokes npm to check exposure has already triggered the attack it was looking for.
Four scan surfaces Covers language package managers (npm, PyPI, Go modules, RubyGems, Composer), MCP AI agent configs, editor extensions (VS Code family), and browser extensions (Chromium + Firefox) — surfaces that typically require separate tools.
Three scan profiles Baseline (routine laptop scan), Project (targeted repo/workspace scan), and Deep (incident‑response sweep across entire home directory).
Open source, Apache 2.0 Available at github.com/perplexityai/bumblebee. Any team can run it with their own threat catalogs.

How Perplexity Uses It Internally

Perplexity runs Bumblebee to protect the developer systems behind its search product, Comet browser, and Computer AI agent. When a new threat emerges — flagged by public disclosures, third‑party intel feeds, or internal research — Perplexity Computer drafts a catalog entry with the ecosystem, package name, and affected versions, then opens a GitHub pull request. A human developer reviews and merges it. Bumblebee then scans all endpoints against the updated catalog and routes findings to the security team.

This pipeline turns a threat signal into an endpoint scan without manual catalog maintenance, according to MarkTechPost. Teams adopting Bumblebee can replicate the same workflow with their own review process.

The Attack Bumblebee Would Have Caught

On May 11, a hacker group tracked by Google as UNC6780 (also known as TeamPCP) planted malicious code into 160+ software packages used by millions of developers. The campaign, which Yahoo Tech reports has been running coordinated software poisoning since at least March 2026, affected packages from Mistral AI, UiPath, and a React tool with 12 million weekly downloads.

“The malicious code fired automatically on install, before anyone noticed anything was wrong,” the Perplexity team said in its.² “A scanner that invokes the package manager to check for infections can trigger those same scripts. You go looking for the worm; the worm runs.”

Bumblebee sidesteps this by reading only metadata — lockfiles and manifests — never executing install scripts, lifecycle hooks, or package manager commands.

MCP Config Scanning Is the Genuinely New Part

Most security tools check software packages and maybe browser extensions. Bumblebee adds a fourth surface that almost nothing else covers: MCP configuration files. These JSON configs tell AI assistants like Claude and Cursor which external services they can connect to — email, databases, calendars, code repositories. A malicious connector slipped into an MCP config could leak credentials or run unauthorized background commands.

Bumblebee parses mcp.json, .mcp.json, claude_desktop_config.json, mcp_config.json, mcp_settings.json, cline_mcp_settings.json, and Gemini CLI settings — without emitting environment values or key names from env blocks. Non‑JSON MCP configs like Codex’s config.toml and Continue’s YAML format are not yet supported in the current v0.1.1 release.

What This Signals About Perplexity’s Strategy

Open‑sourcing internal security tooling is unusual for an AI company in 2026, especially one valued in the tens of billions. Most labs treat their security infrastructure as proprietary competitive advantage. Perplexity’s decision to release Bumblebee under Apache 2.0 signals a bet that developer trust is a product moat — that builders are more likely to adopt Perplexity’s products (Search, Comet, Computer) if they see the company taking supply‑chain security seriously enough to share its tools.

The move also positions Perplexity in a growing conversation about AI agent security. As more developers wire AI assistants into their local environments via MCP, the attack surface expands. Bumblebee addresses that surface directly, years before most security vendors have even acknowledged it exists.

Getting Started

Requirements: Go 1.25 or later. Installation is a single command:

go install github.com/perplexityai/bumblebee@latest
bumblebee selftest

The repository ships with a threat_intel/ directory containing maintained catalogs built from public campaign reporting. Operators control scan cadence through their own scheduling — cron, launchd, systemd, or MDM fleet tooling. Each finding includes a confidence score (high/medium/low), traceable evidence back to the catalog entry, and severity classification.

Sources

1.MarkTechPost(marktechpost.com)
2.announcement(perplexity.ai)

Tags

perplexity bumblebee supply-chain-security open-source developer-tools security-scanner mcp devsec software-supply-chain go-language