"That combination of passive theft and active token requests makes it especially dangerous in cloud build environments, because even short-lived tokens are useful when they're exfiltrated right away," said researchers at OpenSourceMalware.
Malicious updates to Microsoft’s agent-framework: versions 1.26.1 and 1.26.2
Security researchers reported late yesterday and this morning that Microsoft’s official agent-framework package on PyPI was hijacked and used to deliver info-stealing malware. The affected releases are versions 1.26.1 and 1.26.2. According to OX Security, version 1.26.1 was reverted to a safe release, but the attacker later uploaded 1.26.2 through Microsoft’s official PyPI publisher account, indicating the account credentials were still accessible at the time of the report.
How the malicious package operates in compromised environments
Researchers say the infected package does not require being imported to run: installing the malicious package into a Python environment is sufficient for the malware to execute. To maintain the appearance of legitimacy, the malicious installer still pulls the genuine 1.26.0 version of Microsoft’s agent-framework so the library continues functioning while the malware works in the background.
The malware gathers environment variables, credential files, and secrets tied to Azure, Git, Docker, Hugging Face, and other developer services. Captured data is encrypted as a base64-encoded chunk of JSON and transmitted to a webhook.site address hard coded into the malicious code. The infection is non-persistent in the sense that it runs only while the malicious package remains in a site-packages directory; however, its active behavior includes requesting fresh identity tokens from GitHub Actions, Azure DevOps, and Azure, which are then exfiltrated to the same webhook endpoint.

This site is the portfolio.
OSINTSights runs on Cloudflare Workers, D1, R2, and Vectorize, with an AI pipeline on Hetzner ARM. Nubivance designed, built, and operates it. We do the same for clients.
See what we buildImplications for Microsoft’s repository and publisher account controls
Researchers say the incident appears to stem from unauthorized access to an administrator account on Microsoft’s GitHub organization or to its PyPI publisher account. The agent-framework GitHub repository is offline as of writing. Observers also noted that Microsoft’s repo for ONNX Runtime, an AI accelerator, is offline; other Microsoft repositories remain available, and neither Microsoft nor security researchers have linked the ONNX repo’s outage to the malware campaign.
Microsoft did not immediately respond to questions about the incident, and OX Security’s finding that 1.26.2 was uploaded through Microsoft’s official PyPI publisher account raises questions about control and credential handling for publisher accounts.
Why cloud build and CI environments are uniquely exposed
OpenSourceMalware’s assessment highlights a critical operational risk: the malicious package combines passive credential harvesting with active token requests. That makes short-lived, CI-issued tokens — the kind produced by GitHub Actions and Azure DevOps — immediately valuable once exfiltrated. In cloud build environments where ephemeral tokens are used to grant access during builds or deployments, the malware’s ability to request and export fresh identity tokens increases the window of opportunity for an attacker even if the compromised package is present for only a short time.
What this means for technologists, procurement teams, and developers
- Technologists and security teams: If you find agent-framework 1.26.1 or 1.26.2 in your environment, assume compromise and expect to have "some account security work to do," as the reporting put it. The infected package runs without being imported and captures both existing secrets and freshly requested tokens.
- Procurement and enterprise architects: Microsoft’s agent-framework is used to build and deploy AI agents and reportedly sees nearly 200,000 downloads per month. That scale raises supply-chain concerns for organizations that rely on the package in CI/CD pipelines and production build systems.
- Developers and maintainers: Because installation alone can trigger the malware and the package pulls a genuine 1.26.0 to remain operational, casual verification by import or runtime behavior may not reveal compromise. The malicious code was configured to send harvested data to a webhook.site address hard coded into the malware.
The incident leaves immediate, narrow questions that only Microsoft and the researchers can answer from the record: who had access to the PyPI publisher account when 1.26.2 was uploaded, how quickly Microsoft can revoke or rotate those credentials, and whether additional Microsoft repositories were targeted for similar manipulations. Microsoft’s agent-framework repository is offline and Microsoft did not immediately respond to inquiries, even as the company’s conference was under way and some public relations staff were occupied in San Francisco for the event.
For the moment the concrete facts are stark: two packaged releases were weaponized, the attacker used Microsoft’s official PyPI publisher account to publish at least one malicious release per OX Security, the malware exfiltrates both stored and freshly minted credentials (including CI tokens), and the data is sent to a webhook.site destination hard coded into the code. Those details frame both the immediate mitigation work for affected users and the questions about publisher-account controls that Microsoft will need to answer.




