Perplexity launches Bumblebee: How its new read-only dev scanner differs from Chainguard


lapscan-screenshot-2026-05-28-120036

dem10/ iStock / Getty Images Plus via Getty Images

Follow ZDNET: Add us as a preferred source on Google.


ZDNET’s key takeaways

  • Perplexity Bumblebee is an open-source developer security program.
  • Bumblebee doesn’t require AI or a subscription.
  • The program aims to spot problems on programmers’ laptops. 

If you’re a programmer, you’re painfully aware that there’s been a flood of successful malicious attacks on your software supply chain. These attacks include the Axios npm package compromise, the PyPI LiteLLM AI attack, and the CanisterSprawl npm assault

What’s a programmer to do when they can’t even trust the very building blocks of their program? Well, there are several approaches, and the latest comes from Perplexity

According to the AI company, Bumblebee is a “read‑only scanner we use to check developer machines for risky packages, extensions, and AI tool configs during supply‑chain incidents.” The company said in its announcement that the program is one of “the internal tools we use to protect the developer systems behind Perplexity, Comet, and Computer.”

Also: How I got my business emails through spam filters with SPF, DKIM, and DMARC

The security question Bumblebee is built to answer

The tool is built to answer the first question that pops up in your mind after a new supply‑chain advisory: Do any of our programmers have this thing installed? 

Bumblebee runs on MacOS and Linux developer machines and is available now as an open-source Go project. You can plug the tool’s results into whatever security system you’re already using.

Instead of targeting code or runtime behavior, Bumblebee focuses on four specific surfaces. Perplexity claimed existing open‑source tools tend to cover one or two of these surfaces, while Bumblebee can handle all four at once:

  • Language package managers: npm, pnpm, Yarn, Bun, PyPI, Go modules, RubyGems, and Composer
  • AI agent configs: Model Context Protocol (MCP)
  • Editor extensions: VS Code‑family (i.e., VS Code, Cursor, Windsurf, VSCodium)
  • Browser extensions: Chromium‑family (Chrome, Comet, Edge, Brave, Arc) and Firefox

Also: The patching treadmill: Why traditional application security is no longer enough

In other words, this tool is for people running JavaScript/TypeScript, Python, Go, Ruby, and PHP; programmers experimenting with AI MCP configurations; and developers living inside VS Code‑style editors and Chromium‑style browsers.

How Bumblebee integrates into your internal workflow  

Bumblebee is part of a larger internal workflow, which Perplexity outlines as follows:

  1. A threat signal is identified through public disclosures, third‑party intel feeds, or internal research.
  2. Perplexity Computer drafts a catalog update. It enters the signal into a structured entry (ecosystem, name, version), and then opens a GitHub pull request (PR) with source links.
  3. The detection is sent to human review, after which the PR is merged.
  4. Bumblebee runs on endpoints with the updated catalog.
  5. Findings are shared with the security team.

You don’t have to use Perplexity’s JSON catalog; you can now run Bumblebee with your own catalogs and review process. Each detection is “traceable, showing which catalog entry triggered the filing, when it was added, and any evidence,” Perplexity noted.

You can use the open‑source Bumblebee catalog on GitHub. You’ll find it in the threat_intel/ directory, which “holds maintained exposure catalogs built from public threat-intelligence reporting on recent supply-chain campaigns.” Each file in that directory is a catalog in the standard JSON format (schema_version + entries). The README there explains the current catalog list and review guidance. To use the catalogs, you clone the repo and pass that directory to the scanner. For more on that step, see Bumblebee’s Threat Intelligence Exposure Catalogs.

Also: Best VPN services: Expert tested and recommended

Alternatively, you can build your own Bumblebee catalog as a simple JSON file listing exact matches for the risky components you care about, such as ecosystem, package name, and affected versions. Bumblebee then compares local machine inventory against that catalog and flags only exact (ecosystem, name, version) matches, so the catalog is intentionally narrow and deterministic.

The scanner supports three profiles that map pretty cleanly to how developers and security teams think about scope:

  • Baseline profile: Routine scan of standard laptop locations. Teams schedule the scan through their own systems.
  • Project profile: Targeted scan of specific repos or workspaces.
  • Deep profile: Response sweep for active incidents.

Perplexity positions this tool squarely in the “developer surface” tier: Software Bill of Materials (SBOM) and vulnerability scanners handle repositories and build artifacts. Endpoint inventory products cover installed applications. Bumblebee runs on the developer laptop. The key output is: “It tells you whether that machine has a specific package, version, extension, or MCP configuration installed when a supply‑chain advisory lands.”

Read-only avoids risky scans

The company leans hard into “read‑only” as a security property, not just an implementation detail. In their words, “Bumblebee is read‑only. It reads metadata files directly and never lets potentially compromised tooling run, which prevents the scan from becoming a risk.” They added: “Making Bumblebee read‑only helps avoid issues with install‑time code execution.”

Also: 5 ways to fortify your network against the new speed of AI attacks

The post called out npm‑style postinstall attacks directly: “npm packages can carry postinstall scripts that run automatically the moment npm install touches them. That’s how the most recent supply‑chain worms have spread.” The warning for developer‑side scanners is blunt: “A scanner that invokes npm to check for exposure has already triggered the attack it was looking for.”

Bumblebee’s safety guarantees follow from what it refuses to do, said Perplexity:

  • It never runs install scripts or lifecycle hooks.
  • It never runs your package manager.
  • Bumblebee never reads application source files; it reads metadata such as lockfiles, manifests, and installed package metadata.
  • Bumblebee is not an Endpoint Detection and Response (EDR) program.

Framed this way, Bumblebee is not trying to replace endpoint detection tools or build‑time scanners. It’s more of a targeted inventory probe focused on the specific metadata that spots when a particular programmer’s PC is using vulnerable code.

Also: Stopping bugs before they ship: The shift to preventative security

Bumblebee is also not like Chainguard, where the focus is entirely on securing your software supply chain by hardening containers and pipelines rather than developer laptops. The guidance centers on concepts such as minimal, hardened base images, automated rebuilds when vulnerabilities are disclosed, and a policy that blocks non‑compliant artifacts from being shipped.

How Bumblebee compares to Chainguard

Bumblebee lives a step earlier in the lifecycle and a step closer to where developers actually work. Perplexity wrote that “security starts at the local developer surface,” and that “the integrity of our products has to begin further up the supply‑chain than production.” Where Chainguard’s controls surround containers and build outputs, Perplexity said Bumblebee “runs on the developer laptop” and is used “to check developer machines for risky packages, extensions, and AI tool configs during supply‑chain incidents.”

For developers, that approach translates into different touchpoints. Chainguard shows up as base images, policies, and SBOM requirements in your pipelines. Bumblebee is a program your security team runs on your laptop to see which packages, extensions, and MCP configs you currently have installed, and to note which are vulnerable. 

Also: My new favorite Windows app made my PC safer and more reliable – and it’s free

Both approaches have their advantages. Personally, I prefer Chainguard’s approach, which has been expanded to AI tools and code, but I can see how Bumblebee could be useful as well. The tool also has the advantage of being both free and open-source under the Apache 2.0 license. 





Source link

Leave a Reply

Subscribe to Our Newsletter

Get our latest articles delivered straight to your inbox. No spam, we promise.

Recent Reviews


Digital Evidence Has Reshaped Criminal Defense – and the Defense Bar Is Still Catching Up

A decade ago, a felony case file might have run to a few hundred pages of police reports, witness statements, and lab results. Today, that same case can include a full cell phone extraction, hours of body-worn camera footage, surveillance video from multiple cameras, social media exports, license-plate-reader hits, and digital forensic reports running thousands of pages. The substantive law has not changed nearly as fast as the evidence it operates on.

For criminal defense practitioners, the shift is not just about volume. It is about how cases are investigated, how discovery is reviewed, how plea calculations are made, and how trials are tried. A defense lawyer who treats digital evidence as an afterthought — to be skimmed close to trial, with the cell phone dump opened only if something obvious surfaces — is no longer providing competent representation in most serious cases.

The Volume Problem

Modern law enforcement investigations generate digital evidence at a scale that traditional defense workflows were never designed to handle.

A single cell phone extraction using forensic tools commonly used by prosecutors can produce a report tens of thousands of pages long. Multiply that across co-defendants. Add cloud account data subpoenaed from providers. Add body-cam footage from every responding officer, often running an hour or more per officer per incident. Add interview recordings, surveillance video, ALPR records, and any wiretap or pen register data.

The defense lawyer’s obligation is to review all of it — or at least to review it competently enough to identify what matters. Doing that without a workflow is impossible. Cases get lost not because the exonerating evidence was hidden, but because it was buried in the third week of message history nobody had time to read.

The practical response involves a combination of technology and process: e-discovery review platforms scaled for criminal cases, paralegal-level review with defined search protocols, and clear allocation of which categories of evidence the attorney personally reviews versus which are screened first. Firms that handle digital-evidence-heavy cases without that infrastructure tend to discover, late in the process, that something important was missed.

Authentication and Chain of Custody Have Become Central

Volume is half the problem. The other half is that digital evidence is harder to authenticate than the physical evidence it has displaced.

A surveillance video recovered from a business has to be tied to a specific camera, on a specific system, with verified timestamps, with continuous custody from the moment of seizure to the moment of presentation. A cell phone extraction has to be tied to a specific device, performed using a documented forensic process, with hash values demonstrating that the data has not been altered. A social media export has to be authenticated either through the provider’s certification or through circumstantial evidence connecting the account to the defendant.

Each of these chains has potential breaks. Cameras get the wrong time. Forensic extractions get performed with outdated software. Social media accounts get used by people other than the registered user. Defense counsel who understands the technical underpinnings of how evidence was collected can identify gaps that opposing counsel may have assumed were settled.

Federal procedure in particular has evolved around these issues. Practitioners working in federal court should be familiar with the Federal Rules of Evidence governing authentication and the best-evidence rule, both of which apply to electronic records in ways that often surprise lawyers more accustomed to paper-era practice.

Discovery Obligations and the Brady Problem

The growth of digital evidence has also complicated the prosecution’s obligations under Brady and its progeny, which require disclosure of material exculpatory and impeachment evidence to the defense.

When the relevant evidence universe was a few hundred pages, prosecutors could reasonably review the file and identify Brady material. When the universe is a hundred thousand pages of cell phone data and dozens of hours of video, identifying what is exculpatory becomes a much harder problem — and not always a problem prosecutors solve well. Defense counsel cannot rely on the prosecution to flag what the defense will find useful. The defense has to find it themselves, which loops back to the volume problem.

Courts have been inconsistent in how they handle Brady obligations in the digital age. Some jurisdictions require prosecutors to provide searchable, organized productions; others permit document dumps that effectively shift the search burden to the defense. The practical implication is that defense lawyers in serious cases must budget significantly more time for discovery review than would have been required even a few years ago, and must do so on schedules that prosecutors and courts often have not adjusted to reflect the new reality.

How Digital Evidence Changes Plea Negotiations

Plea negotiations have always been driven by each side’s assessment of trial risk. Digital evidence has changed both sides of that calculation.

For the prosecution, video and digital records often appear to lock in factual elements that previously turned on witness credibility. A clear video of an alleged assault, or a series of incriminating messages, can shift a case from a battle of testimony into a battle of interpretation. Prosecutors evaluating cases with strong digital evidence often offer less, because they perceive their trial position as stronger.

For the defense, the same evidence frequently contains nuance that changes how a jury would actually receive it. Body-cam footage that the prosecution thinks is damning often shows context that supports the defense theory. Cell phone messages read in full rather than excerpted often tell a different story. The defense lawyer who has actually watched the video and read the messages — rather than relying on the prosecution’s characterization — is often in a meaningfully stronger negotiating position than the case file would initially suggest.

This is part of why pretrial preparation has become more decisive. The cases that resolve favorably are usually the cases where the defense did the digital evidence work early enough to see what was actually there, rather than what the police reports said was there. Resources from the California Courts and the State Bar of California outline the procedural framework within which this work has to happen, but the framework alone does not produce results — sustained attention to the evidence does.

What Effective Defense Looks Like Now

Competent criminal defense in 2026 looks different than it did even five years ago. The lawyers who get the best outcomes for clients tend to share a few characteristics: they take digital evidence seriously from intake forward, they have the infrastructure to review it at scale, they understand the technical questions well enough to challenge authentication where appropriate, and they treat plea calculations as something to be made after the evidence has been examined rather than after the police reports have been read.

For people facing serious charges in California, the practical implication is that the choice of counsel matters more, not less, in the digital evidence era. A firm like Angelo Reyes Law, built around trial-ready preparation rather than volume-driven plea processing, reflects what effective representation tends to look like in cases where the evidence record is large and where the difference between a good and a poor outcome turns on what defense counsel actually finds in the file.

The volume of evidence will keep growing. Defense practice has to keep up.



Source link