Socket’s cover photo
Socket

Socket

Computer and Network Security

Socket is the #1 software supply chain security platform. Next-gen SCA + SBOM + 0-day prevention. LOVED BY DEVELOPERS.

About us

Socket is a cybersecurity platform that protects companies from software supply chain attacks. Companies use Socket to protect their software applications and critical services from malware and security threats originating in open source code.

Industry
Computer and Network Security
Company size
51-200 employees
Headquarters
San Francisco
Type
Privately Held
Founded
2020
Specialties
Software, Security, Software supply chain, Open source software, Application Security, Cybersecurity, and Software Composition Analysis (SCA)

Locations

Employees at Socket

Updates

  • Socket reposted this

    The attack didn't start with code. It started with a collaboration email. The maintainer of one of the most widely used HTTP libraries in JavaScript, the de facto standard, got a note from what looked like a legitimate company. They invited him to a Slack channel. For nearly a month they worked with him: real conversations, real ideas, real help for his project. Then a Microsoft Teams call cut out mid-meeting and a prompt appeared saying Teams was out of date, click this file to update. The file was malware. Several of us at Socket received the same lure. I can't take credit for not falling for it. I never saw the email. One of the best phishing defenses, it turns out, is being too busy to read your inbox. Not a strategy I would recommend. :) This is what supply chain attacks look like now: patient, targeted, and aimed at the human rather than the code. Compromise one maintainer and you reach every application that depends on the library, increasingly including the AI agents and coding tools that pull it in by default. You can't train your way out of a month-long social engineering campaign. You have to assume compromise will happen and catch malicious code at the moment it enters your dependencies. That's what we're building at Socket. From my conversation with Anne Dwane on the Village Global podcast.

  • Socket reposted this

    Five frontier LLMs. Nearly 200,000 coding prompts. All five hallucinated the same nonexistent package names. After review by PyPI Security and Socket, 53 of those names remain open for anyone to register on PyPI or npm. Independent researcher Aleksandr Churilov replicated the USENIX Security 2025 package-hallucination methodology against the 2026 model cohort: Claude Sonnet 4.6, Claude Haiku 4.5, GPT-5.4-mini, Gemini 2.5 Pro, and DeepSeek V3.2. The findings: 🔺 199,845 code-generation responses across ~40,000 Python and JavaScript prompts 🔺 Overall hallucination rates of 4.62% to 6.10%. The spread across models narrowed from 16.5 percentage points in the 2025 cohort to 1.48, but no model beat the earlier study's best result of ~3.6% (GPT-4 Turbo). 🔺 Python hallucination rates exceeded JavaScript across all five models, reversing the earlier study's finding 🔺 127 hallucinated names appeared in output from every model: 109 on PyPI, 18 on npm 🔺 53 remain registrable after review: 41 on PyPI, 12 on npm. Public examples: css-color-stop, dns-sd, dom-ains. The intersection is the story. This is not five models making five different mistakes. Models from Anthropic, OpenAI, Google, and DeepSeek converge on identical fake package names, likely via shared training material or shared ecosystem conventions. One malicious registration could surface in AI-generated code across all of them. That is what makes slopsquatting practical. To be clear: these are potential targets, not an active attack. There is no evidence any of the 53 names have been registered maliciously. Socket reviewed all 18 npm candidates and removed six false positives. The four highest-volume "hallucinations" were valid Ember.js imports (@ember/service, @ember/object, @ember/routing, @ember/controller) that resolve through ember-source but 404 on a registry lookup. Absence from a registry does not always mean a hallucination. PyPI Security's prohibited-name list and normalization protections had already blocked 68 of the 109 Python candidates. If your team ships AI-generated code: treat every AI-suggested dependency as unverified until you have confirmed the name, publisher, and history. Lockfiles and pinning prevent drift, but they do not make an unfamiliar package safe the first time it enters your project. Review AI-suggested dependencies the way you would review a dependency proposed by an unknown external contributor.

    • No alternative text description for this image

Similar pages

Browse jobs