How A PyPI Hijack Targeted Credentials Through Microsoft’s AI Framework
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Two releases of Microsoft’s official agent-framework package on PyPI, versions 1.26.1 and 1.26.2, contained credential-stealing code on November 17, 2026. PyPI quarantined both, but anyone who installed either version should treat the environment as exposed, rebuild from a clean base and rotate credentials it could access.

Two releases of Microsoft’s official Agent Framework Python package on PyPI contained a credential stealer on November 17, according to the GitHub Advisory Database and security researchers. The versions, 1.26.1 and 1.26.2, could run malicious code whenever Python started in an affected environment, while the framework continued to work normally.

Before you orderOffer from Amazon

Get everyday essentials delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

PyPI records show that version 1.26.1 was published at 20:35 UTC, followed by version 1.26.2 at 21:08 UTC. StepSecurity says PyPI quarantined the first release at 21:06 UTC, about 31 minutes after publication, and quarantined the second at 21:16 UTC, eight minutes after it appeared. The current PyPI release listed in the source material is the clean version 1.26.0, published November 12.

The GitHub Advisory Database says both affected versions contained a credential stealer absent from 1.26.0, and rates the incident critical. The advisory says the malicious code could run without a user importing agent_framework. That means an installation could appear to work as expected even as code in the Python environment ran at interpreter startup. The package is the one Microsoft’s installation instructions point to, rather than an unrelated lookalike.

The package’s scope matters because it is an umbrella for Microsoft’s open-source toolkit for building AI agents in Python and coordinating them in workflows. The security firm Wiz says the package receives almost 200,000 downloads a month; the source material does not give a measurement period or a comparison baseline for that figure. Security firms including Socket, StepSecurity, Wiz, JFrog and Endor Labs analyzed the releases. The source material identifies credentials, cloud keys and CI tokens as targets, and says the second release included renamed files and additional capabilities.

At a glance
reportWhen: Published November 17, 2026; advisories…
The developmentTwo malicious releases of Microsoft’s agent-framework Python package were published to PyPI on November 17 and later quarantined.
Bar chart titled 50 stored requests: who sent them? Based on Wiz's analysis of the requests stored by the attacker's free collection service: 29 from sandboxes, vendor scans and tests; 17 carrying real victim data; 4 from the attacker's own machines. Highlighted: behind the 17 requests, 3 victim cases, which Wiz counts as unique, successful thefts of victim data. Note: only what the free service stored, at most 50 requests; no install count has been published
What the attacker's collection service stored, per Wiz Research (researchers' claim, updated November 18, 2026). Chart: Thorsten Meyer AI.

Credentials at Risk Across Python Workflows

The incident could affect more than a developer’s local machine. The package may have been installed on laptops, build servers, continuous integration pipelines or in container images, including through automatic updates. If malicious code ran in an environment that could access secrets, those credentials may need attention even if no visible problem appeared in the software.

Quarantine prevents later installation from PyPI, but it does not by itself remove a copy already present in a local environment, cache, mirror or built image. StepSecurity reported that the second release appeared just over two minutes after the first was quarantined. It said quarantine had not immediately revoked the attacker’s ability to publish. JFrog similarly cautioned that quarantining a version does not revoke a compromised publishing credential or prevent a later upload. Those details show why package removal and credential rotation address different parts of the exposure.

For organizations, the practical concern is whether either version entered an environment with access to sensitive credentials. The advisory specifically calls for rotating credentials the affected environment could reach, including CI and GitHub Actions secrets. The source material does not establish how many installations occurred during the exposure windows or whether credentials were successfully taken.

Products Worth Considering

Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera – Midnight

Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera – Midnight

As an affiliate, we earn on qualifying purchases.

Two Releases After a Clean Build

The last clean release identified in the advisory is agent-framework 1.26.0, published November 12. The next release, 1.26.1, appeared five days later as a wheel-only upload, according to the advisory. Endor Labs says releases it examined from 1.21.0 through 1.26.0 had also included source code. The difference was one of the details researchers reported while examining the package history.

After PyPI quarantined 1.26.1 at 21:06 UTC on November 17, version 1.26.2 appeared at 21:08 UTC, according to StepSecurity’s timeline. The second version used renamed files and, according to the source material, added capabilities targeting GitHub Actions tokens. PyPI quarantined it at 21:16 UTC. The OpenSSF malware database listed the case as MAL-2026-20101 at 22:12 UTC on November 17; the GitHub Advisory Database published GHSA-764m-3ppc-5m73 at 00:31 UTC on November 18.

Wiz said it reported its findings to Microsoft on November 17. The source material attributes the quarantine times to StepSecurity and the affected-version assessment to the GitHub advisory. These are separate records of the incident, and their publication does not establish the number of affected users or the extent of any credential theft.

“A credential stealer that is not present in the last clean release, 1.26.0.”

— GitHub Advisory Database

Exposure and Investigation Remain Open

The available reports do not say how many people or organizations installed either affected version, how many copies remain in caches or images, or whether credentials were successfully exfiltrated. They also do not identify which specific credentials were accessed in any confirmed victim environment. A package’s ability to seek credentials is not proof that it obtained or misused them.

The source material reports Microsoft’s brief statement to The Register but says no Microsoft security advisory had been found as of its publication on November 19. It does not provide a detailed account of the publishing-access compromise, its duration, or whether access has since been revoked. Those points remain unconfirmed in the supplied reporting.

Rebuild and Rotate Exposed Secrets

Users and administrators should check package records, lockfiles, build logs and deployed images for agent-framework 1.26.1 or 1.26.2. The advisory’s recommended response, as summarized in the source material, is to treat an environment that installed either version as exposed, rebuild it from a clean base using 1.26.0, and rotate every credential the environment could reach, including CI and GitHub Actions secrets.

Teams should also check whether affected package copies entered internal caches, mirrors or container images, since removing a release from PyPI does not remove those copies. Further details may come from Microsoft or updated security advisories. The source material does not give a timeline for Microsoft’s investigation or a date for any additional public report.

Key Questions

Which versions of agent-framework were affected?

The GitHub Advisory Database identifies versions 1.26.1 and 1.26.2 as containing the credential stealer. The previous release, 1.26.0, is identified as the last clean version in the supplied advisory summary.

Could the malicious code run if I did not import the framework?

According to the GitHub advisory, the code could run without importing agent_framework, when Python started in an affected environment. Whether it accessed credentials depends on that environment and what it could reach.

What should an organization do after installing an affected version?

Treat the environment as exposed, rebuild it from a clean base on version 1.26.0, and rotate credentials it could access. The advisory’s guidance, as provided in the source material, specifically includes CI and GitHub Actions secrets.

Does PyPI quarantine prove no credentials were stolen?

No. Quarantine stops a release from being installed through PyPI, but does not remove copies already in environments, caches, mirrors or images. The supplied reports do not establish whether credentials were successfully taken.

Source: Thorsten Meyer AI

You May Also Like

Is Your Data Safe? The Emotional Impact of AI Security

AIThis post was created with the assistance of artificial intelligence (AI). As…

Warning! AI Security: The Critical Factor in Protecting Your Personal Data

AIThis post was created with the assistance of artificial intelligence (AI). After…

AI Security: The Game-Changer in Modern Cyber Defense

AIThis post was created with the assistance of artificial intelligence (AI). As…

The Alarm That Caught Hackers Targeting A Blood Bank’s Plasma Freezers

France’s blood service says alarms caught altered freezer settings at three sites. Plasma was moved, with no reported product loss or health impact.