🔍 Read the full analysis: Who Has A Part In Responding To The Microsoft AI Agent Framework PyPI Hijack? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Malicious versions 1.26.1 and 1.26.2 of Microsoft’s agent-framework Python package were available on PyPI on November 17, 2026, and contained credential-stealing code, according to the GitHub Advisory Database and security researchers. PyPI quarantined both releases; affected users are advised to treat environments that installed them as exposed, rebuild from a clean version and rotate accessible credentials. The number of affected installations and whether credentials were used remain unknown.
Two malicious releases of Microsoft’s agent-framework Python package were briefly available on PyPI on November 17, and contained code that sought developer and cloud credentials, according to public security advisories. PyPI quarantined both versions, but developers and organizations that installed either release must treat the affected environments as exposed and investigate credentials the code could reach.
The affected versions were 1.26.1 and 1.26.2. The GitHub Advisory Database says both included a credential stealer absent from the last clean release, 1.26.0, and rates the incident critical. According to the advisory, the code could run when Python started in an affected environment, even if the user did not import the agent-framework package. That meant routine activity such as running a build or a linter could trigger it after installation.
StepSecurity’s timeline puts the first upload at 20:35 UTC on November 17 and PyPI’s quarantine of it at 21:06 UTC. A second release appeared at 21:08 UTC and was quarantined at 21:16 UTC. The source describes the initial exposure as lasting about 31 minutes. Both versions have since been removed from PyPI; the source reports that 1.26.0, released November 12, was the current version there at the time of reporting.
The advisory says the malware sought credentials in environment variables and developer-tool files, and could request tokens from services including Azure and Azure DevOps. Version 1.26.2 also added GitHub Actions identity-token access, according to the source. The recommended response is to rebuild any affected environment from a clean base using version 1.26.0 and rotate every credential the environment could access, including CI and GitHub Actions secrets. Teams should investigate related accounts and systems rather than assuming that removing the package alone resolves possible credential exposure.
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
- Clean 1.26.0 publishedStill the current version on PyPI
- Malicious 1.26.1 appearsFramework works normally, a credential stealer runs whenever Python starts
- PyPI quarantines 1.26.1After about 31 minutes
- Attacker uploads 1.26.2Renamed files, also targets GitHub Actions tokens
- PyPI quarantines 1.26.2Eight minutes later
- Malware databases list itMAL-2026-20101 (OpenSSF), GHSA-764m-3ppc-5m73 (GitHub)
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 Build Systems Are Exposed
The incident matters beyond users who deliberately chose to install unfamiliar software. The source describes agent-framework as the package referenced by Microsoft’s installation instructions, and says security firm Wiz estimates it receives almost 200,000 downloads a month. That figure is a monthly download estimate, not a count of affected installations; the number of users who installed either malicious release has not been established.
Developer laptops and automated build systems can hold credentials that grant access to source code, cloud services, package registries and deployment pipelines. If a malicious package can read those credentials or obtain tokens, an attacker may be able to use them beyond the machine where the package ran. The advisory and researchers describe potential access, not confirmed downstream account compromise. Organizations need to assess what each affected environment could reach and whether those credentials were used.
The response also cannot rest with package users alone. Microsoft must investigate how the publishing channel was compromised and restore confidence in its releases; PyPI must contain malicious uploads and review platform-side controls; and organizations must rebuild environments and manage exposed secrets. Security researchers and advisory maintainers contribute by analyzing the package, publishing findings and helping defenders identify the scope of the risk.
Popular right now

Vitamix 5200 Professional-Grade Blender with 64 oz Container, Black

TCL 65-Inch R7D Series 4K UHD QLED Roku Smart TV

Philips 5500 Series Fully Automatic Espresso Machine with LatteGo Milk System
As an affiliate, we earn on qualifying purchases.
How the Two Releases Appeared
The package on PyPI is named agent-framework and is described in the source as an umbrella package that pulls in framework code and integrations. Microsoft Agent Framework is an open-source toolkit for building AI agents in Python and coordinating them in workflows. The legitimate 1.26.0 release was published on November 12, according to the source’s account of PyPI records.
After the first malicious release was quarantined, an attacker published 1.26.2 about two minutes later. StepSecurity says the second upload followed the initial quarantine by 2 minutes and 23 seconds. The researchers’ explanation, as reported in the source, is that quarantine of a release did not itself revoke the publishing credential or prevent a new upload. That distinction makes the second release relevant to the response: removing a bad version and securing the account or publishing process are separate tasks.
Several security firms, including Socket, StepSecurity, Wiz, JFrog and Endor Labs, analyzed the releases, while the OpenSSF malware database and GitHub Advisory Database published records. The source says Wiz reported its findings to Microsoft on November 17. Microsoft told The Register: “We have made some repositories private while we investigate malicious content.” The source reported no Microsoft security advisory at publication time.
“We have made some repositories private while we investigate malicious content.”
— Microsoft, in a statement to The Register
Victim Count and Credential Use
The available source material does not give a confirmed number of affected installations, nor does it establish how many credentials were collected or whether stolen credentials were used. The monthly download estimate cannot answer those questions because it does not identify downloads of the two malicious versions or distinguish installations from other package activity.
It is also not clear from the material how the attacker obtained publishing access, whether any other Microsoft repositories or packages were affected, or how many environments had the malicious releases installed through automated dependency updates. Microsoft said it was investigating malicious content but did not provide a detailed incident account in the statement cited. Security researchers reported no remote-code-loading second stage in the inspected code; that finding is limited to the code they examined and does not establish the full extent of any compromise.
Actions Across the Supply Chain
Organizations should first determine whether laptops, containers, build agents or CI jobs installed 1.26.1 or 1.26.2, including through automatic updates. Following the advisory’s guidance, they should rebuild affected environments from a clean base on 1.26.0 and rotate credentials the environment could access. Security teams should also review relevant cloud, source-control and CI activity for suspicious use; the source does not specify a confirmed list of compromised accounts.
Microsoft’s investigation needs to clarify the publishing compromise, affected repositories and measures taken to prevent further unauthorized releases. PyPI’s quarantine contained the two versions, but the incident also puts attention on how publishing access is secured and what controls follow a malicious upload. Further advisories or incident updates may establish the scale of exposure. Until then, affected organizations should base remediation on whether the versions were installed and what secrets were available to those environments, not on the short time the releases were listed.
Key Questions
Which package versions were affected?
The affected PyPI releases were agent-framework 1.26.1 and 1.26.2. The source identifies 1.26.0, released November 12, as the last clean version cited in the advisory.
What should an organization do if it installed one of the releases?
Treat the environment as exposed, rebuild it from a clean base using 1.26.0, and rotate credentials it could access, including CI and GitHub Actions secrets. Review relevant systems for suspicious credential use as part of the investigation.
Did the malicious code require importing the framework?
No. The GitHub Advisory Database says the code could run when Python started in an affected environment, without importing agent_framework.
How many people or organizations were affected?
The source does not report a confirmed victim count. Wiz’s estimate of almost 200,000 downloads a month describes package activity generally, not downloads or installations of the two malicious releases.
Who is responsible for responding?
Users and organizations must identify affected environments, rebuild them and rotate accessible credentials. Microsoft must investigate its publishing channel and repositories; PyPI’s role includes containing malicious releases and addressing platform-side controls. Security researchers and advisory maintainers provide analysis and warnings to support those responses.
Source: ThorstenMeyerAI.com
