Incident sheet · /security

Filed
Read
3 min · 598 words
Severity
Routine

Microsoft Agent Framework malware steals cloud credentials on PyPI

Two malicious Microsoft Agent Framework releases on PyPI were built to steal cloud and developer credentials. Versions 1.26.1 and 1.26.2 ran during Python startup and were quarantined.

Microsoft’s Agent Framework PyPI package compromised to steal credentials
FeedMicrosoft’s Agent Framework PyPI package compromised to steal credentials — kuszapro / pixabay

Event log

  1. Malicious versions 1.26.1 and 1.26.2 were published on November 17, 2026.
  2. Both releases ran a credential stealer during normal Python startup, even without an import.
  3. The package recorded roughly 200,000 monthly downloads on PyPI.
  4. Version 1.26.2 appeared about 2.5 minutes after 1.26.1 and was quarantined about eight minutes after publication.
  5. Affected teams should rebuild with 1.26.0 and rotate potentially exposed credentials immediately.

By Olivia Hart

Microsoft's official Agent Framework package was compromised on November 17, 2026, with malicious versions 1.26.1 and 1.26.2 published to PyPI and later quarantined.

The danger for an enterprise developer sat inside ordinary Python startup. Both releases carried a credential stealer that executed automatically in affected environments, even when Microsoft’s Agent Framework package was not imported.

Microsoft Agent Framework is the company’s open-source toolkit for building AI agents and coordinating them in workflows. The package draws roughly 200,000 monthly downloads on PyPI, giving a malicious update a broad enterprise distribution channel.

The reported 200,000 monthly downloads measure package use, not confirmed infections; no victim count or financial-loss estimate has been published. BleepingComputer said it had contacted Microsoft to learn "what the point of breach was" and reported: "We have not received a response by publication time."

How did the two malicious versions work?

Version 1.26.1 harvested secrets from environment variables, local configuration files and other running processes wherever permissions allowed. Its targets included:

  • Azure
  • AWS
  • GitHub
  • Docker
  • Hugging Face
  • OpenAI
  • Other services

The release also attempted to obtain Azure managed identity access tokens and Azure DevOps OIDC tokens. If successful, the tokens could give attackers access to additional cloud resources.

The requests targeted identity artifacts that can carry permissions into Azure and Azure DevOps. That makes the exposure a cloud-access problem, not merely a local workstation cleanup.

Version 1.26.2 added GitHub Actions OIDC token requests and searches of GitHub runner directories, extending the collection beyond the services named in version 1.26.1.

Both malicious versions retained a dependency on the legitimate agent-framework-core 1.26.0 package. The framework continued to function normally while the malware ran in the background, leaving developers with an apparently healthy package and a compromised environment.

Socket, Step Security and Ox Security discovered and documented the attack.

What did researchers find inside the releases?

According to Socket and Step Security, version 1.26.1 used agent_framework_meta.pth to execute the credential stealer in agent_framework_meta/_telemetry.py during Python startup.

For version 1.26.2, the attacker restored the original package initializer and moved the malicious code into af_runtime_diag.pth and af_runtime_diag.py.

Both versions exfiltrated collected data to the same attacker-controlled endpoint: https://webhook[.]site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb. Investigators should search for that exact webhook URL path in egress logs, distinguishing blocked requests from transfers that completed.

The shared destination gives responders a single exact indicator across both releases. It is also the more relevant search term in environments where the package was never imported but Python startup still triggered the payload.

PyPI quarantined the first release, but the attacker published version 1.26.2 roughly 2.5 minutes later. The second release was quarantined about eight minutes afterward. That sequence shows the attacker still had publishing access after the initial quarantine, although how it obtained that access remains unclear.

What should developers do now?

Developers who installed either malicious version should act in this order:

  • Isolate the affected environment.
  • Avoid running Python or pip until the environment is secured.
  • Rebuild from a known-clean environment using the safe 1.26.0 release.
  • Rotate immediately any potentially exposed credentials within the compromised environment’s reach.
  • Search egress logs for the exact webhook path and separate blocked requests from completed transfers.

For enterprises, containment, clean rebuild and credential rotation must run together; log review then establishes whether the attacker’s collection stopped at the network boundary or data left the environment.

via bleepingcomputer.com (Source)

Filed under

  • microsoft-agent-framework
  • pypi
  • software-supply-chain
  • cloud-credential-theft
  • malware
Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

News editor covering money, funding and market shifts around enterprise information technology and the leaders who run at CIO Monitor.

207 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›