A poisoned AI development package could steal cloud credentials before a developer even used it.
Two malicious releases of Microsoft’s agent-framework Python package appeared on PyPI on Nov. 17, targeting credentials accessible to developer machines and automated build systems. Versions 1.26.1 and 1.26.2 have since been removed, according to a GitHub security advisory published Nov. 18.
For teams that installed either version, removing the package is only the beginning. The urgent question is which credentials the affected environment could reach and whether those credentials remain usable.
Two malicious releases, 33 minutes apart
Endor Labs’ analysis places the first upload at 20:35 UTC and the second at 21:08 UTC. The researchers found malicious additions to the distributed package rather than a replacement of the framework’s core functionality.
The distinction matters: Software can keep working while its installation introduces a separate threat. Endor said the releases retained their dependency on agent-framework-core version 1.26.0.
Machine-learning tools have faced supply-chain attacks before, including a compromised PyTorch dependency that TechRepublic covered. That earlier incident is context, not evidence of a shared attacker or technique. The sources reviewed do not establish how access to publishing was obtained in this case.
Why launching Python could trigger the theft
Endor found that both releases added executable .pth startup hooks. Python processes executable lines in these files during normal interpreter startup, allowing the malicious helper to run without the developer having to import Agent Framework.
The second release moved its payload into different files while leaving the package’s initialization file identical to the clean version, Endor reported. Reviewing only that file could therefore miss the malicious additions.
The GitHub advisory describes the collection of environment variables and credential files, alongside Azure and build-system tokens. The second release also targeted GitHub Actions tokens. This is a risk to affected development and cloud environments, not evidence that every Microsoft account or Windows device was compromised.
Must-read security coverage
- UK Police Convicts Pair in £5.5 Billion Bitcoin Launder Case
- Blackpoint Cyber vs. Arctic Wolf: Which MDR Solution is Right for You?
- How GitHub Is Securing the Software Supply Chain
- 8 Best Enterprise Password Managers
Check exposure, then revoke access
The advisory identifies 1.26.1 and 1.26.2 as malicious and 1.26.0 as the last clean release. It advises treating environments that installed either malicious version as exposed and rotating credentials reachable from them.
Security teams should coordinate that response across development, cloud operations, and identity owners:
- Identify affected installations. Review dependency records, build logs, and deployed images for the two versions. Check automated pipelines as well as developer workstations.
- Contain before rebuilding. Preserve relevant evidence and keep suspect environments from receiving replacement secrets. Rebuild from verified dependencies rather than assuming package removal reverses credential theft.
- Scope the credentials. Establish which accounts and services the environment could access, then revoke or rotate exposed credentials from a trusted system.
- Review subsequent activity. Look for unexpected cloud actions, repository changes, and pipeline activity associated with those identities.
Reduce the access an infected build can inherit
The broader lesson is permission scope. TechRepublic’s coverage of enterprise AI agent security gaps examines why access controls deserve as much scrutiny as new capabilities. This incident reaches the development tools behind those systems, making the build environment part of that review.
For IT leaders, a useful follow-up is to map which development jobs can reach production services. Separate test credentials from production access, document who owns each service identity, and avoid giving experimental projects the same privileges as established applications. These are risk-reduction measures, not substitutes for investigating an affected installation.
TechRepublic’s reporting on long-lived cloud credentials also explains why reusable secrets can extend an attacker’s opportunity. Review credential lifetime and permissions together: Reducing either can limit exposure, but neither makes a compromised environment safe to keep using.
Removing a malicious release from a registry prevents some future installations. It does not invalidate secrets already taken. For affected teams, recovery ends with restored trust in the environment and its credentials, not simply a clean dependency list.
Related reading: Learn why AI agents create insider-security blind spots and where identity governance can fall short.