Microsoft’s agent-framework package on PyPI has reportedly been compromised to distribute a credential stealer. According to The Hacker News RSS report published on November 18, 2026, the malicious releases are 1.26.1 and 1.26.2, published 33 minutes apart on November 17. The report says the payload is designed to execute whenever Python starts, making this a concern for developer workstations, build runners, and other environments holding cloud credentials.

The full source article could not be retrieved for this advisory. The incident details below are therefore limited to the available RSS summary; the response steps are general defensive recommendations, not additional confirmed findings about this attack.

What is reported—and what remains unconfirmed

The Hacker News attributes reporting on the compromise to Aikido, Endor Labs, OX Security, SafeDep, Socket, StepSecurity, and Wiz. The available summary identifies the two malicious package versions and describes a credential-stealing payload with Python-startup execution.

It does not establish the initial compromise method, provide indicators of compromise, identify a verified clean replacement release, or explain how startup execution is implemented. Do not assume that another version is safe merely because it is not named here. Check current maintainer and security-researcher advisories before selecting a replacement.

Identify exposed environments

Prioritize systems that installed or updated agent-framework, especially automated builds and developer environments with access to production services. Review lockfiles, dependency manifests, package-install logs, artifact inventories, and container build records for 1.26.1 or 1.26.2.

Because the reported payload runs at Python startup, avoid launching Python or Python-based package-management commands inside a suspected environment just to inspect its version. Where practical, examine files and logs using trusted tooling outside that environment. Finding a dependency declaration alone does not prove execution; installation records and runtime evidence help determine actual exposure.

Include cached dependencies and reusable build images in the investigation. A removed upstream release can still exist in an internal mirror or previously built artifact.

Contain exposure and protect credentials

If a malicious release was installed, isolate the affected workstation or runner from sensitive services and pause pipelines that reuse it. Preserve relevant installation logs, build records, and security telemetry before rebuilding.

Treat credentials accessible to that environment as potentially exposed, rather than assuming every account in the organization was compromised. From a separate trusted system, revoke or rotate affected cloud keys, service-principal credentials, deployment tokens, and other relevant secrets. Invalidate active sessions where supported and review cloud audit logs for unexpected authentication, privilege changes, or resource creation.

Removing the package alone is not a sufficient recovery plan: it cannot invalidate stolen credentials. Rebuild affected environments from trusted sources, validate the chosen dependencies, and only then reintroduce replacement secrets.

Reduce repeat exposure

Temporarily block the reported malicious versions in dependency policies and internal repositories. Use reviewed dependency updates, controlled artifact promotion, and short-lived, least-privilege credentials for build jobs. Version pinning improves reproducibility, but pinning a compromised release simply makes the compromise reproducible; hashes likewise verify an artifact’s identity, not its safety.

The immediate priority is to establish whether the named releases reached your systems, contain confirmed or credible exposure, and secure credentials before restoring normal operations.

Source: The Hacker News report on the Microsoft Agent Framework PyPI compromise.