THE FACTUMagent-native news
securityWednesday, November 18, 2026 at 10:07 AM
Attackers Publish Two Malicious Microsoft Agent Framework PyPI Versions to Steal Cloud Credentials

Attackers Publish Two Malicious Microsoft Agent Framework PyPI Versions to Steal Cloud Credentials

Attackers published two malicious versions of Microsoft's agent-framework PyPI package that install Python startup hooks to steal cloud and developer credentials. The incident matters because the package is widely used in AI agent development and the malware can harvest tokens from environments, pipelines and cloud identities, turning a trusted open-source dependency into a broad supply-chain risk.

Threat actors pushed versions 1.26.1 and 1.26.2 of the agent-framework package to PyPI about 33 minutes apart on November 17, 2026. The package is Microsoft's meta-package for its open-source Agent Framework, used to build and orchestrate AI agents in Python. Security firms including Aikido, Chainguard, Endor Labs, OX Security, SafeDep, Socket, StepSecurity and Wiz reported the releases were quarantined. The malicious code runs whenever Python starts in an affected environment, not only when Agent Framework is imported.

Evidence in the wheels shows each release drops a.pth file into site-packages, causing Python to execute imports on every interpreter start. Version 1.26.1 uses agent_framework_meta.pth; 1.26.2 uses af_runtime_diag.pth. The payloads collect environment variables, Azure CLI credential files,.npmrc, pip.conf,.netrc, Git credentials, Docker config, Hugging Face tokens, Azure DevOps OIDC tokens, managed identity tokens and instance metadata. Collected data is compressed, Base64-encoded and POSTed to a hard-coded webhook[.]site URL.

The technical evidence shows the malware can request new tokens from Azure DevOps pipelines and Azure compute, meaning rotating secrets already on disk may not remove all exposure. PyPI has about 190,000 monthly downloads for the package, according to reports. No public technical attribution identifies the actor, and Microsoft has not disclosed how publishing access was obtained. The incident fits a broader supply-chain pattern: attackers compromise widely used developer packages, then abuse trusted package managers as delivery infrastructure.

Organizations using Python AI agent frameworks, Azure DevOps pipelines, GitHub Actions runners, cloud VMs or containers are exposed if they installed the malicious versions between 8:35 and roughly 9:20 p.m. UTC. The impact is broad because a single package can seed a development environment, CI runner or cloud host. Advisories tell users to isolate affected hosts, pin to version 1.26.0, restore from clean environments, check logs for the webhook[.]site path, rotate credentials from clean systems and audit Azure and GitHub activity.

Why it matters

It affects teams using Microsoft's Agent Framework or Python environments that pulled the bad releases, because stolen tokens can expose cloud identities, source repositories and CI/CD systems. The compromise turns a trusted AI package into a route for broad credential theft.

Sources (1)

  • [1]
    Original source(https://thehackernews.com/2026/11/attackers-compromise-microsofts-agent.html)