UTC

Supply-chain campaigns against the AI stack.

14 campaigns that put malicious code into AI and agent packages, extensions and skills: which releases were affected, what to do if you installed one, and the reports they come from.

  1. 2026-11-17

    Microsoft agent-framework PyPI compromise

    Two releases of agent-framework, the Python package of Microsoft Agent Framework, were uploaded to PyPI without being built from the official repository. Both ran a credential stealer at every Python start, aimed at Azure, Azure DevOps and GitHub Actions tokens. It was disclosed on 2026-11-17, so details may still change.

    agent-framework

  2. 2026-10-21

    LiteLLM release pipeline compromise (1.107.0)

    LiteLLM's own release pipeline was hijacked with stolen developer credentials and published a malicious 1.107.0 to PyPI, Docker Hub and GHCR. The release collected cloud, AI provider and proxy credentials and loaded remote code. It is separate from the March 2026 incident.

    litellm

  3. 2026-06-17 · Sapphire Sleet, according to Microsoft Threat Intelligence

    @mastra npm scope compromise

    A maintainer's npm account with rights across the @mastra scope was taken over and used to publish new versions of more than 140 Mastra packages. Each added one dependency, easy-day-js, a lookalike of dayjs whose malicious release installed a remote-access implant aimed at credentials, browser data and wallets.

    easy-day-js, mastra, @mastra/core, create-mastra and 3 more

  4. 2026-05-19 · TeamPCP, according to Wiz and the PyPA advisory (likely)

    Microsoft durabletask PyPI compromise

    Three releases of durabletask, Microsoft's Durable Task Python client, were published to PyPI with a stolen publishing token, outside the normal pipeline. On import they fetched a credential-theft framework aimed at cloud, Kubernetes, Vault and password-manager secrets.

    durabletask

  5. 2026-05-12 · TeamPCP, according to OSV (Google Open Source Security) and StepSecurity

    guardrails-ai 0.10.1 malicious release

    With an employee's GitHub token, the attacker reached the project's publishing credentials and released a malicious guardrails-ai 0.10.1 to PyPI. It carried the self-spreading credential stealer of the May 2026 Mini Shai-Hulud wave, and PyPI quarantined it within about two hours.

    guardrails-ai

  6. 2026-05-11

    Mistral AI SDK compromise

    Part of the May 2026 Mini Shai-Hulud wave. A compromised developer device led to malicious Mistral SDK releases on PyPI (mistralai 2.4.6, which has no matching tag or pipeline run) and npm. The PyPI release fetched a credential harvester on Linux; Mistral describes the npm payload as non-functional.

    mistralai, @mistralai/mistralai, @mistralai/mistralai-azure, @mistralai/mistralai-gcp

  7. 2026-04-30

    PyTorch Lightning PyPI compromise

    Two releases of lightning on PyPI ran injected code on import that started an infostealer aimed at credentials, CI secrets and cloud secret stores. It tried to spread with stolen tokens and left hooks in Claude Code and VS Code settings to run again. The maintainers have not published a root cause.

    lightning

  8. 2026-04-29 · TeamPCP for the May wave, according to StepSecurity and OSV (Google Open Source Security)

    Mini Shai-Hulud worm (2026)

    A self-spreading worm across npm and PyPI that steals publishing tokens and uses them to release infected versions of the victims' own packages. After an SAP wave in late April 2026, the May wave hit AI packages, including the Mistral SDKs, guardrails-ai and several MCP servers, alongside 42 TanStack packages.

    guardrails-ai, mistralai, @mistralai/mistralai, @mistralai/mistralai-azure and 6 more

  9. 2026-04-22 · Disputed: JFrog linked it to TeamPCP, which denied involvement

    Xinference PyPI compromise

    A malicious pull request took secrets from the project's CI runner, and three trojanized xinference releases followed on PyPI. They ran on import and collected credentials, environment variables, SSH keys, cloud tokens and wallet data.

    xinference

  10. 2026-03-24 · TeamPCP, according to JFrog Security Research

    LiteLLM PyPI compromise (1.82.7 and 1.82.8)

    Two litellm releases were uploaded straight to PyPI, outside the project's pipeline, with a credential stealer that ran on import (and, in 1.82.8, at every Python start). The publishing credentials are believed to have leaked through a compromised security scanner in LiteLLM's CI.

    litellm

  11. 2026-01-27

    ClawHavoc: malicious OpenClaw skills

    Large-scale poisoning of ClawHub, the skill registry of the OpenClaw agent. Hundreds of skills posed as useful tools and told users to install a fake prerequisite that delivered infostealers on macOS and trojans on Windows.

  12. 2026-01-06

    Fake n8n community nodes

    Malicious npm packages posed as n8n community nodes, most as a Google Ads integration. Once installed, they read the OAuth credentials stored in n8n when a workflow ran and sent them to the attacker.

    n8n-nodes-hfgjf-irtuinvcm-lasdqewriit, n8n-nodes-gg-udhasudsh-hgjkhg-official, n8n-nodes-ggdv-hdfvcnnje-uyrokvbkl, n8n-nodes-vbmkajdsa-uehfitvv-ueqjhhhksdlkkmz and 5 more

  13. 2025-10-17

    GlassWorm

    Malware in VS Code-compatible extensions, first on Open VSX and later on the VS Code Marketplace, published mostly with stolen publisher tokens. It hid its loader in invisible characters, collected developer credentials and wallets, and used stolen tokens to publish more. Later waves added macOS and a malicious MCP server on npm.

    codejoy.codejoy-vscode-extension, l-igh-t.vscode-theme-seti-folder, kleinesfilmroellchen.serenity-dsl-syntaxhighlight, JScearcy.rust-doc-viewer and 15 more

  14. 2025-09-15

    Shai-Hulud npm worm (2025)

    A self-spreading npm worm that stole publishing tokens and released infected versions of the victims' packages, in September 2025 and again in November 2025 ("The Second Coming"). The second wave hit MCP servers and AI SDKs from Postman, Zapier, Browserbase, PostHog, mcp-use and Voiceflow, among hundreds of other packages.

    @postman/postman-mcp-server, @postman/postman-mcp-cli, @zapier/mcp-integration, @zapier/ai-actions and 15 more

Each campaign is written up by hand from the reports it links to and checked on the date it shows. ShieldMap names an actor only when a primary source does. Report a correction to [email protected].