Two malicious versions of Microsoft’s agent-framework Python package were uploaded to PyPI on November 17, carrying code designed to steal cloud credentials and CI/CD secrets. The affected releases, 1.26.1 and 1.26.2, have been removed from the public index, but copies may remain in caches, internal mirrors and container images.
The payload uses a Python startup hook, so an affected environment can run the stealer even when an application never imports Agent Framework. Teams that installed either release should isolate affected systems, investigate access and rotate reachable credentials from a clean environment. GitHub’s malware advisory identifies 1.26.0 and earlier as unaffected by these malicious releases; it lists no patched version.
Where to look first
Search dependency lockfiles, installation and build logs, software inventories, virtual environments, container images and internal package mirrors for agent-framework==1.26.1 or agent-framework==1.26.2. Include short-lived CI runners and images built during the exposure window. A clean current requirements file does not rule out an earlier installation or a cached wheel.
Wiz researchers recommend pinning to the known-clean 1.26.0 while investigating. Verify the artifact’s source and integrity in a clean build environment, and remove the malicious wheels from caches and mirrors. Do not assume that an unpinned upgrade or removal from PyPI has repaired systems that already installed the package.
If an environment may be affected, avoid starting Python or running pip there during initial triage: either can trigger the startup hook. Isolate the system in line with your incident-response procedures, preserve relevant evidence, and inspect its files and logs from a trusted environment. Rebuild affected environments and images from known-clean sources rather than relying on an in-place uninstall.
What runs, and what it tries to steal
Agent Framework is used to build and orchestrate AI agents. The affected agent-framework distribution is a top-level package; its legitimate core dependency continued to work, making the malicious addition less obvious. According to JFrog’s analysis, the wheel places an executable .pth file in Python’s site-packages directory. A normal interpreter startup processes the hook before user code runs. Installing the wheel does not itself execute the hook, but a subsequent Python process in that environment can.
For version 1.26.1, look for agent_framework_meta.pth and agent_framework_meta/_telemetry.py. Version 1.26.2 uses af_runtime_diag.pth and af_runtime_diag.py. Searching only for the first version’s filenames can miss the second. These indicators are useful leads, not a substitute for a broader investigation: absence of a file after cleanup does not establish that the payload never ran.
The stealer searches environment variables and credential files, and attempts to read other processes’ environments where permissions allow. GitHub’s advisory lists Azure CLI, Docker, Git and Hugging Face credentials among its targets. It also attempts to obtain Azure managed-identity and Azure DevOps identity tokens; the second release adds GitHub Actions token collection. The collected data is compressed and sent to an attacker-controlled webhook.
Contain exposure beyond the package
Treat credentials reachable by affected hosts and runners as potentially exposed, even if there is no proof that every credential was transmitted. From a separate trusted system, revoke and replace relevant cloud keys, application secrets, GitHub and package-registry tokens, and CI/CD credentials. Review managed identities and federated trust relationships, and examine cloud audit logs for unusual token use, new credentials or unexpected changes to permissions.
Review proxy, DNS and outbound-traffic logs for the specific indicators in the Wiz investigation. A successful outbound request can help establish exposure, but missing logs or a lack of a visible request should not by itself clear a system. Retain logs and affected artifacts for incident responders before disposal.
What remains uncertain
Wiz’s investigation reports evidence suggesting that an attacker obtained a certificate and private key used in Microsoft’s release process. How those credentials were obtained remains unknown. JFrog found that the malicious wheels were repacked outside the normal build pipeline; that does not by itself establish the route used to publish them. Neither conclusion should be read as proof that Microsoft’s source repository was modified.
The two releases appeared roughly 33 minutes apart. Their removal limits new exposure from the public index, but the practical risk persists anywhere an affected artifact was copied or deployed. Track the GitHub advisory and your vendors’ incident guidance for updates as the investigation develops.