Two releases of Microsoft's Agent Framework Python package turned an official software download into a credential-stealing threat. The malicious code could execute whenever Python was started in an affected environment, without developers having to import the framework.
Versions 1.26.1 and 1.26.2 appeared on the Python Package Index (PyPI) on Nov. 17, targeting cloud credentials, developer secrets, and CI/CD tokens. Although both releases were quarantined within an hour, organizations that downloaded them could still have compromised environments.
Security teams must identify affected installations, investigate potential data theft, and replace exposed credentials, rather than assume that removing the packages eliminated the threat.
Two malicious releases exploited Python startup behavior
According to the PyPI release history, versions 1.26.1 and 1.26.2 were published on Nov. 17 before being quarantined. Version 1.26.0, released Nov. 12, remains the latest unaffected version.
The first malicious release appeared at 20:35 UTC, followed by the second at 21:08 UTC, approximately 33 minutes later.
Researchers at StepSecurity found that both releases contained executable .pth startup hooks. These files allowed the credential stealer to run when Python initialized the affected environment, even if an application never explicitly imported Agent Framework.
The malicious packages retained their legitimate dependencies, allowing the framework to continue functioning while the stealer operated.
Version 1.26.1 used files named agent_framework_meta.pth and _telemetry.py. The second release changed the malicious components to af_runtime_diag.pth and af_runtime_diag.py, making detection based solely on the original filenames insufficient.
Malware targeted Azure, GitHub, and developer credentials
The payload collected environment variables, configuration files, and credentials accessible through other running processes.
Researchers identified attempts to obtain Azure managed-identity tokens and Azure DevOps OpenID Connect (OIDC) tokens. The malware also searched for Git credentials, Docker configuration data, package-manager credentials, and authentication information used by development environments.
Version 1.26.2 expanded the attack by targeting GitHub Actions OIDC tokens and additional runner directories.
These capabilities created risks beyond individual developer machines. A compromised CI/CD runner with access to cloud identities or deployment credentials could expose resources connected to an organization's build and deployment infrastructure.
The malware compressed the collected information, encoded it, and attempted to transmit it to an attacker-controlled endpoint hosted through webhook.site.
Researchers confirm limited successful exfiltration
An updated Wiz investigation examined 50 requests captured by the attacker's webhook endpoint.
Researchers attributed four requests to attacker activity and 29 to testing or security scanning. The remaining 17 requests represented three distinct cases of successful victim data exfiltration.
The findings provide evidence that the attack resulted in the actual exposure of credentials. However, they do not establish how many organizations installed the malicious versions or whether stolen credentials were subsequently used to access additional systems.
Wiz also reported evidence suggesting that the attacker compromised a client certificate and private key associated with Microsoft's Engineering System Release Pipeline. The researchers had not established how those credentials were obtained.
The incident follows other software supply chain attacks targeting development environments, including a GitHub account compromise affecting thousands of repositories, illustrating the risks posed by stolen developer and automation credentials.
How organizations should respond to the compromised packages
The GitHub security advisory identifies versions 1.26.1 and 1.26.2 as malicious and recommends treating affected environments and accessible credentials as exposed.
Security teams should prioritize three actions:
- Identify affected installations: Examine dependency lockfiles, CI/CD logs, developer environments, container images, and package caches for either of the malicious versions. Check for the associated .pth startup files and suspicious outbound connections.
- Contain and rebuild compromised environments: Isolate affected systems before running Python or pip, as starting the interpreter could re-execute the payload. Rebuild from a known-clean environment, pin Agent Framework to version 1.26.0, and remove cached malicious packages.
- Rotate credentials and investigate cloud access: Replace potentially exposed secrets from a clean system. Review Azure identity activity, GitHub Actions OIDC permissions, service connections, and cloud audit logs for unauthorized access.
Removing the malicious package alone is insufficient if the stealer has already transmitted credentials.
Security teams should also distinguish attempted network connections from confirmed data transfers. Blocked requests do not necessarily indicate successful exfiltration, but the presence of previously executed malicious code still warrants investigation.
Package removal does not eliminate existing exposure
The attack demonstrates how quickly a compromised software release can reach developer machines and automated build systems.
Quarantining the malicious versions prevents ordinary new downloads from PyPI, but existing installations, internal mirrors, and cached package files can remain a source of exposure.
The incident is particularly relevant to organizations adopting AI development frameworks that interact with cloud services and automated deployments. As previous research into AI software supply chain threats has highlighted, trusted development dependencies can provide attackers with access to sensitive infrastructure.
For affected organizations, the immediate priority is establishing whether the malicious versions were executed and which credentials were accessible. Restoring a clean package is only part of recovery; identifying and invalidating exposed access is equally important.
Also read: For a related incident, read how the LiteLLM supply chain attack exposed credentials across AI ecosystems.





