Every week, our malware detection systems scan thousands of new and updated packages across public registries like npm and PyPI. This week was no exception.
We confirmed over 90 malicious packages between June 26 and July 3, 2026, across npm and PyPI, with several campaigns continuing from previous weeks and new ones emerging.
The nolimit-agent campaign continued publishing new versions (1.0.306, 1.0.315, 1.0.316), extending the Microsoft 365 device-code phishing framework we documented in detail in our DeviceDoor report. The anthropic-toolkit cluster was the dominant new campaign, with 20 versions confirmed on June 30 alone — directly targeting packages associated with Anthropic’s developer tooling. The cursed-modules family published over 15 versions between July 1 and July 2 across both standard and inflated version numbers (999.x), following the dependency confusion playbook. The date-fns-lite cluster (5 versions, July 2) continued the pattern of impersonating legitimate, widely-used utility packages.
The AI tooling targeting pattern established last week with ollama-helpers and openai-agents-helpers extended into this period with ai-explain, ai-sdk-helpers, and @langgraphjs/toolkit, all confirmed between June 28 and June 30. The @szc-ft/mcp-szcd-client package (versions 0.38.0 and 0.39.0, confirmed July 2) introduced a new pattern we are tracking as SkillLeak: a credential decryptor hidden inside an MCP skill rather than an install hook, invisible to scanners that stop at postinstall. We published a full SkillLeak analysis here.
This weekly snapshot is part of our ongoing Malicious Code Digest, where we validate new threats and provide actionable intelligence to help DevSecOps teams protect their pipelines before damage occurs. Let’s break down what we found this week and why it matters.
90+ Packages. One Week. The Pipeline Is the Target.
This week’s digest reflects a shift in attacker focus, not just volume, but precision. Sustained version flooding, AI tooling clusters, MCP-layer credential theft, and dependency confusion attacks against internal monorepo namespaces. The campaigns are automated and continuous. A weekly scan is not a defense.
Xygeni Early Malware Warning monitors npm, PyPI, and other registries in real time, flagging threats at the moment of publication, before they reach a build, before an AI agent installs them autonomously, and before a SkillLeak-style payload has a chance to execute. When anthropic-toolkit publishes 20 versions in a single day or cursed-modules floods npm with version 999.x across two days, detection that runs after the fact is already too late.
Xygeni’s Open Source Security platform gives DevSecOps teams the real-time detection and prioritization needed to stay ahead of coordinated supply chain pressure, so your pipelines stay clean without slowing your teams down.
