Two malicious releases of Microsoft's Agent Framework package on PyPI exposed developers and build systems to a credential stealer, according to analyses published November 18 by StepSecurity and Elastic Security Labs. Versions 1.26.1 and 1.26.2 of the Python package were uploaded on November 17 and later quarantined by PyPI. The code targeted cloud and continuous integration credentials in environments that installed either release.

Agent Framework is Microsoft's open-source toolkit for building AI agents. The affected agent-framework package is a meta-package that pulls in other components, according to Elastic. GitHub's malware advisory identifies only versions 1.26.1 and 1.26.2 as malicious and says the last clean release was 1.26.0. The available evidence does not establish how the attacker gained publishing access or how many installations were affected.

Both malicious wheels added Python .pth startup hooks, StepSecurity found. Python processes these files when it starts in an affected environment, allowing the stealer to run without an application importing Agent Framework. The first release used agent_framework_meta.pth and a telemetry module; the second used files named af_runtime_diag.pth and af_runtime_diag.py. The legitimate framework dependency remained in place, which could let normal package functions continue.

StepSecurity and Elastic reported that the payload searched environment variables, credential files and other accessible process environments for secrets. It also attempted to obtain Azure managed identity, Azure DevOps and GitHub Actions identity tokens. The researchers traced attempted exfiltration to a webhook.site endpoint. A matching network request alone would not prove that data was successfully transferred, StepSecurity cautioned.

StepSecurity's timeline places the first upload at 20:35 UTC on November 17 and the second at 21:08 UTC, after PyPI had quarantined the first release. Both were quarantined within about 41 minutes of the initial upload. GitHub's advisory says the malicious versions have since been removed from PyPI. No public source reviewed here establishes a confirmed downstream breach or the total number of exposed users.

Organizations that installed either version should identify the affected machines and build jobs, isolate them before launching Python tools, and rebuild from a clean environment, according to StepSecurity's response guidance. Teams should rotate credentials the payload could access and review egress logs for the specific endpoint identified in the research. Pinning to the verified clean 1.26.0 release and checking cached packages and container images can prevent a quarantined wheel from being reused.