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

🔍 Read the full analysis: The Credential Risks Behind The Microsoft AI Agent Framework PyPI Hijack on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Two PyPI releases of Microsoft’s agent-framework package, versions 1.26.1 and 1.26.2, contained credential-stealing code on November 17, 2026. PyPI quarantined both, but researchers say the package could run the code when Python started, even if the framework was not imported. Anyone who installed either version should treat the environment as exposed, rebuild it from a clean base and rotate accessible credentials.

Two releases of Microsoft’s Agent Framework Python package on PyPI contained code that searched for and exfiltrated credentials, according to security advisories and researchers. Versions 1.26.1 and 1.26.2 were published on November 17 and quarantined within minutes, but the code could run when Python started in an affected environment, without the framework being imported.

The GitHub Advisory Database identifies the affected package as agent-framework and says the malicious code was absent from the prior clean release, version 1.26.0. The advisory rates the incident critical. PyPI’s current version is reported as 1.26.0, released November 12. The affected package is the one Microsoft’s installation instructions direct users to, rather than an unrelated package using a similar name.

According to the timeline compiled by StepSecurity, version 1.26.1 appeared at 20:35 UTC on November 17 and PyPI quarantined it at 21:06. Version 1.26.2 was uploaded at 21:08 and quarantined at 21:16. The first release was available for about 31 minutes; the second for about eight minutes, based on those timestamps. StepSecurity says the second upload came shortly after the first was quarantined, indicating that blocking one release did not by itself stop another upload.

The advisory says a startup hook caused the code to run when Python launched in the affected environment, even if a user did not import agent-framework. Researchers say it sought credentials in environment variables, developer-tool and container-registry files, and cloud and CI environments. Version 1.26.2 added collection of GitHub Actions identity tokens, according to the source material. The real framework remained functional, which could make the compromise less apparent to users.

At a glance
reportWhen: Incident occurred November 17, 2026; in…
The developmentResearchers documented credential-stealing code in two releases of Microsoft’s Python Agent Framework package on PyPI, which were published and quarantined on November 17.

agent-framework on PyPI · UTC

Two poisoned releases, one evening

The official package of Microsoft’s Agent Framework, step by step. Blue: what happened. Green: the clean version.

November 12 to 18, 2026

  1. Clean 1.26.0 publishedStill the current version on PyPI
  2. Malicious 1.26.1 appearsFramework works normally, a credential stealer runs whenever Python starts
  3. PyPI quarantines 1.26.1After about 31 minutes
  4. Attacker uploads 1.26.2Renamed files, also targets GitHub Actions tokens
  5. PyPI quarantines 1.26.2Eight minutes later
  6. Malware databases list itMAL-2026-20101 (OpenSSF), GHSA-764m-3ppc-5m73 (GitHub)
Quarantine is not revocation. The second release came 2 minutes 23 seconds after the first was pulled: the attacker still had publishing access. Copies already in caches, mirrors and images stay there.

Sources: GitHub Advisory Database GHSA-764m-3ppc-5m73, Nov 18, 2026; OSV MAL-2026-20101, Nov 17, 2026; PyPI project data, checked Nov 19, 2026; StepSecurity analysis (quarantine times).

Graphic: Powered by Thorsten Meyer AI

Why Cloud Credentials Were Exposed

The potential impact reaches beyond a developer’s local computer. A machine or build environment can hold cloud keys, source-control tokens, package credentials and AI service access. If stolen credentials remained valid, an attacker could potentially use them to access systems or services permitted by those credentials. That is a risk assessment, not confirmation that any particular account was accessed.

The package’s startup behavior also matters for response: a person may have installed the release without explicitly using the framework and still run Python afterward. CI pipelines, build servers and containers may have started Python during routine testing or deployment. The GitHub advisory advises treating an environment that installed either affected version as exposed, rebuilding from a clean base on 1.26.0 and rotating credentials it could reach, including CI and GitHub Actions secrets.

Researchers describe the incident as a supply-chain compromise because malicious code was distributed through the public registry in a package associated with Microsoft’s project. The source material says Wiz estimated the package receives almost 200,000 downloads a month; that figure is a reported download estimate, not a count of affected installations or confirmed victims.

Products Worth Considering

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite

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.

How the PyPI Releases Unfolded

Microsoft’s Agent Framework is an open-source toolkit for building Python-based AI agents and coordinating them in workflows. The agent-framework package is described as an umbrella package that brings in the framework’s core code and integrations. PyPI is a public registry commonly used to distribute Python packages.

In the sequence reported by the advisory and security researchers, Microsoft published clean version 1.26.0 on November 12. Endor Labs said releases it examined from 1.21.0 through 1.26.0 had included source code, while 1.26.1 was uploaded as a wheel only. After PyPI quarantined 1.26.1, another release followed before the second was also quarantined. The OpenSSF malware database listed the case as MAL-2026-20101; the GitHub Advisory Database published GHSA-764m-3ppc-5m73.

Security firms including Socket, StepSecurity, Wiz, JFrog and Endor Labs analyzed the releases, according to the source material. Wiz said it promptly reported its findings to Microsoft on November 17. The source also reports that PyPI removed both affected releases, while researchers urged users to take action even though the exposure window was short.

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

— GitHub Advisory Database

Victim Counts and Access Still Unknown

The source material does not provide a confirmed count of users or environments that installed versions 1.26.1 or 1.26.2, or a count of credentials actually stolen. A short availability window does not establish how many installations occurred, particularly where updates were automated. The reported monthly download estimate is not a measurement of incident victims.

It is also unclear whether any exfiltrated credentials were used to access cloud accounts, repositories, pipelines or other services. The source material says StepSecurity found no remote-code-loading second stage in the code it inspected and reports that the malware was not persistent beyond the installed package. Those findings do not establish whether stolen credentials were used elsewhere, nor do they settle the total scope of the incident.

Microsoft’s reported statement says repositories were made private while the company investigated malicious content. The source material provides no fuller Microsoft advisory or public account of how the publishing access was compromised, which repositories were affected, or whether additional releases or systems were involved.

Rebuild, Rotate and Await Findings

Organizations should check package records, lockfiles, build logs and container histories for installations of agent-framework 1.26.1 or 1.26.2. The advisory’s recommended response is to treat an environment that installed either release as exposed, rebuild it from a clean base using version 1.26.0, and rotate every credential the environment could access. Teams should include CI secrets and credentials available to cloud workloads in that review.

Further updates depend on findings from Microsoft and the researchers, including whether the publishing credential or account was compromised, how many environments installed the affected releases and whether investigators found use of stolen credentials. The source material reports the releases as quarantined and 1.26.0 as the current PyPI version; readers should check current advisories and package records for later changes.

Key Questions

Which package versions were affected?

The affected PyPI releases were agent-framework 1.26.1 and 1.26.2. The source material identifies 1.26.0, released November 12, as the last clean version.

Did users have to import the framework for the code to run?

No, according to the GitHub Advisory Database. It says the code used a startup hook and could run when Python started in an affected environment, without importing agent-framework.

What should an organization do if it installed an affected version?

The advisory recommends treating the environment as exposed, rebuilding from a clean base on version 1.26.0 and rotating credentials it could access, including CI and GitHub Actions secrets. Teams should consult the advisory for details suited to their systems.

How many victims were affected?

The source material gives no confirmed victim or installation count. Wiz’s reported estimate of almost 200,000 monthly downloads is not a count of installations of the two malicious versions or of compromised environments.

Has Microsoft explained how the releases were published?

The source material reports a brief statement to The Register saying Microsoft had made some repositories private while investigating malicious content. It does not provide a public explanation of how publishing access was compromised or the investigation’s final findings.

Source: ThorstenMeyerAI.com

You May Also Like

Robert Wille (Munich Quantum Software Company): Quantum Software Needs Design Automation, Not Just Algorithms

Robert Wille argues quantum computing needs production-ready design automation and interoperable software, not just new algorithms.

AnySpeechRecognizer And AnyDecisionModel: A New Option For Apple AI

Hugging Face announced two Swift packages for choosing speech and decision models through shared APIs, including options designed to run on device.

Small Businesses And AI: Turning Ideas Into Action

An OpenAI page is titled “Helping small businesses put AI to work,” but available details do not confirm an announcement, product or program.

Beyond AI Productivity: The Effort Of Reviewing Its Results

Research finds AI can speed some work but slow other tasks when people must check and revise its output.