Home / Blog / Microsoft Agent Framework Releases Poisoned With Credential…
Tech News

Microsoft Agent Framework Releases Poisoned With Credential Stealer

The package serves as Microsoft's foundation for building multi-agent AI systems and records roughly 190,000 monthly downloads across the Python ecosystem.

By Dillip Chowdary • Nov 18, 2026 • Source: SecurityWeek

Microsoft Agent Framework Releases Poisoned With Credential Stealer
Image: SecurityWeek

At a glance

Product
Microsoft Agent Framework
Compromised versions
1.26.1 and 1.26.2
Clean version
1.26.0
Attack date
November 17
Monthly downloads
Approximately 190,000
Affected ecosystems
Azure, Azure DevOps, GitHub Actions

Attackers distributed two compromised releases of Microsoft’s Agent Framework Python package through the Python Package Index in a November 17 supply chain attack, according to SecurityWeek's report. The malicious versions targeted developer workstations, container images, and continuous integration workflows, attempting to harvest cloud authentication tokens, environment variables, and repository secrets.

This report details the mechanics of the compromised distributions, the credential theft payloads, the scope of exposure across Python environments, and the remediation steps required for affected development teams. It serves software engineers, DevOps teams, and security administrators who run the package to build AI agents and multi-agent systems across cloud infrastructure.

01Malicious package versions and PyPI quarantine

Attackers published two back-to-back malicious versions of the agent-framework package to PyPI, designated as versions 1.26.1 and 1.26.2. The releases appeared approximately 33 minutes apart on November 17 before registry administrators and package maintainers intervened. The package serves as Microsoft's foundation for building multi-agent AI systems and records roughly 190,000 monthly downloads across the Python ecosystem.

Security teams quarantined both versions on PyPI shortly after discovery to halt automatic distribution to downstream systems. A Microsoft Agent Framework maintainer confirmed in a GitHub discussion that the compromised files were removed from PyPI while an active investigation into the intrusion proceeded. Clean deployments require explicitly locking systems to the preceding verified package.

Note: Version 1.26.0 remains the latest clean, verified release on PyPI, and versions 1.26.1 and 1.26.2 should be treated as entirely compromised.

02Python startup exploitation and payload mechanics

Independent security assessments conducted by OX Security, Wiz, Endor Labs, and SafeDep identified an automatic execution vector embedded directly into the distributions. Rather than requiring developers to import the library inside application code, the malware attaches to the Python startup mechanism. The payload triggers automatically whenever the Python interpreter initializes inside an infected environment.

The attackers retained the original library components alongside the malicious routines to avoid immediate detection during routine unit tests or agent tasks. Because legitimate application calls continued to function normally, infected workloads showed no obvious runtime crashes. This allowed the package to persist unnoticed within background workflows while extracting credentials.

The credential stealer targets running process memories, configuration files, and system environment variables. In cloud deployments, the malicious code attempts to pull Azure managed identity tokens and harvest Azure DevOps authentication tokens. The subsequent 1.26.2 release expanded this harvesting routine to actively collect GitHub Actions secrets.

03Package release metadata and telemetry comparison

The compromised packages lacked the standard source archives that typically accompany Microsoft's verified distribution pipeline. Security researchers reviewing PyPI release histories found clear anomalies between the clean build history and the two rogue uploads:

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Package ReleaseDistribution ArtifactsSource Repository CodeCredential Stealer ScopeRegistry Status
Version 1.26.0Built and source distributionsVerified in main repositoryNoneActive clean release
Version 1.26.1Built packages onlyAbsent from main repositoryAzure and Azure DevOpsQuarantined on PyPI
Version 1.26.2Built packages onlyAbsent from main repositoryAzure, DevOps, GitHub ActionsQuarantined on PyPI

Security researchers at Wiz confirmed that neither malicious release corresponded to pull requests, tags, or commits within Microsoft's main source repository. The rogue uploads bypassed the verified build pipeline entirely, indicating that the artifacts were published directly to PyPI from an external system.

SafeDep identified a new maintainer account named vibiret that was registered on PyPI on the day of the attack. The account possessed publishing permissions for the agent-framework project along with five associated packages, though its direct involvement remains under review. Additionally, Microsoft's main Agent Framework repository and related projects, including ONNX Runtime, became inaccessible on GitHub during the incident investigation.

04Version pinning and environment verification

Development teams must immediately verify package manifests, clean local caches, and lock dependency trees to known-good artifacts. The following terminal commands check installed packages, enforce the clean version, and pin containerized workloads:

Command
pip show agent-framework
Command
pip install --upgrade "agent-framework==1.26.0"
Command
pip cache purge
Command
echo "agent-framework==1.26.0" > requirements.txt

Automated build systems should enforce strict hash checking and dependency pinning to prevent build agents from resolving quarantined versions. Teams using dependency management tools should verify lockfiles to ensure that transient dependencies do not pull release candidates or poisoned patch versions.

05Environment remediation and token rotation

Organizations that ran either compromised package must isolate affected nodes and perform a complete credential rotation across their cloud and source control infrastructure. Security administrators should execute the following sequence:

  1. Audit dependencies by scanning container registries, virtual environments, and continuous delivery pipelines for agent-framework versions 1.26.1 or 1.26.2.
  2. Quarantine runtime hosts and terminate active developer instances, runner nodes, and container pods where the affected versions were executed.
  3. Rotate exposed secrets including all Azure managed identity configurations, Azure DevOps personal access tokens, GitHub Actions repository secrets, and environment variables.
  4. Inspect cloud audit logs across Azure and GitHub Actions to identify anomalous resource access, unauthorized token exchanges, or unexpected API calls originating from affected hosts.
  5. Rebuild clean environments from base images using explicit dependency constraints pinned to version 1.26.0.

Frequently asked questions

Which versions of Microsoft Agent Framework were compromised?

Versions 1.26.1 and 1.26.2 were published maliciously to PyPI and contain credential-stealing code. Version 1.26.0 remains the latest verified clean release.

How does the malware execute without an explicit import?

The malware exploits Python's startup mechanism to execute automatically whenever the interpreter begins running, even if the application does not import the library directly.

What specific credentials does the malware steal?

It extracts environment variables, configuration files, Azure managed identity tokens, Azure DevOps authentication tokens, and GitHub Actions secrets from the environment.

Were the malicious changes merged into Microsoft's source repository?

No, researchers found no corresponding code or releases in Microsoft's main repository, and the malicious uploads included only built packages without source distributions.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →