Payments are experiencing issues due to temporary restrictions in Russia. If your payment does not go through, please submit a support request.Our support team is available 24/7 — we are always here to help with hosting and server issues.We are now accepting requests for dedicated server rental and colocation services in our data center.Reminder: we recommend enabling backups for additional data protection.A new VPS/VDS lineup with NVMe storage and improved performance is now available.Maintenance work on some servers has been completed. All services are operating normally.
Article4 min readViews1

Malicious versions of agent-framework removed from PyPI: what to check in projects and CI

Researchers discovered credential theft in agent-framework versions 1.26.1 and 1.26.2. Why removing the package is not enough and how to safely verify your environment.

Comments 0

Cardboard box with a torn coral-colored paper seal at the lid joint
In this article

Malicious versions 1.26.1 and 1.26.2 of the agent-framework package—a meta-package for the Microsoft Agent Framework for Python—have been removed from PyPI. According to initial reports from Wiz and SafeDep, these versions steal credentials from development or build environments. If these versions were included in a project, container, or CI pipeline, stop using the suspicious environment immediately and check all secrets it could access. Removing the package does not undo code that has already executed or prevent potential data leakage.

What happened and who is affected

According to Wiz, version 1.26.1 was published on November 17, 2026, at 20:35 UTC, and version 1.26.2 at 21:08 UTC. This corresponds to 23:35 MSK on November 17 and 00:08 MSK on November 18, respectively. By 21:30 UTC, both versions were no longer available in PyPI. At the time of our check, public PyPI metadata listed 1.26.0 as the current version and contained no release files for 1.26.1 or 1.26.2.

The risk applies to environments where these two releases were installed, including developer workstations, server containers, and CI runners. Merely using Python, Azure, or other Microsoft products does not prove infection. The reason for gaining access to the package publication remains unconfirmed, and the number of affected installations is also unknown.

Why the absence of an import does not provide protection

Malicious releases add an autostart file to the Python environment. Its code executes when the interpreter starts, even if the application does not import agent-framework. Therefore, a successful build or lack of calls to the library does not rule out code execution. Checking a suspicious environment by running its Python or pip is dangerous: this may re-trigger data theft.

Researchers described the collection of environment variables, credential files, and cloud tokens. Access to package registries, repositories, and cloud resources may be compromised. Managed Azure identities and federated CI access are particularly critical: version 1.26.2 also requests GitHub Actions tokens when the job has the necessary permissions. This poses a risk not only of leaking stored passwords.

Safe order for verification and recovery

  • Suspend jobs using the suspicious environment and isolate affected runners. Preserve CI logs, images, and dependency lock files for analysis before cleanup.
  • From a clean system, verify which versions were actually installed by checking lock files, installation logs, and the contents of saved images. Do not run Python from the environment being checked; the absence of two versions in the current PyPI does not prove they were not installed previously.
  • Recreate the environment from verified sources and packages. Compare the entire dependency graph and hashes, not just the top-level agent-framework. Before resuming work, prevent unverified new releases from entering the build.
  • From a clean system, revoke or replace environment secrets and sessions. Check cloud access and activity logs since installation, managed identity permissions, and trust in federated CI tokens. Changing a single password does not close all these access vectors.

Wiz and SafeDep refer to the release of agent-framework 1.26.0 as the last clean version. However, SafeDep separately highlights dependent projects that may update independently. Therefore, pinning only this version number does not replace checking the entire package set, and rolling back within an already compromised environment does not replace recreating it and reviewing access.

Installing a malicious version requires checking for potential secret exposure, but it does not prove that an attacker successfully accessed all related services. Actual damage depends on environment permissions and subsequent data usage. We did not run samples or perform infrastructure checks on General iT or client systems for this material.

Information verified on November 18, 2026, at 12:06 MSK. Details on the compromise and the status of releases may be updated.

Discussion 0

Share your experience and ask questions. Comments without links appear after editorial review.

No comments yet. Start the discussion.