Your lockfile can be clean today while an image built three days ago still contains a compromised dependency. That matters for teams responding to the Microsoft Agent Framework incident. On 17 November, two malicious releases of the legitimate agent-framework package, 1.26.1 and 1.26.2, appeared on PyPI. They targeted credentials available to developer machines and build jobs.
What the package analyses establish
Wiz and Elastic independently examined the releases. Both found code that can run when Python starts in an affected environment, even if the application never imports Agent Framework. Their analyses describe attempts to collect secrets from environment variables and local credential files and to obtain cloud and CI credentials. A build that finished successfully can still be part of the exposure window.
PyPI removed both versions. At verification on 21 November, its listing showed 1.26.0, the last release the advisory identifies as unaffected by this incident. How the attacker obtained publishing access, who was behind the uploads and which credentials left individual organisations remain unknown from the public evidence.
Check what actually ran
Start with the affected package and version in build logs, dependency inventories, image manifests and deployment records. Look back from the 17 November uploads through the last use of any image, runner or cache that could contain either release. Check developer environments and private mirrors too. A downstream job can reuse an artifact well after a release disappears from the public registry.
- Installed in a runner, image or workstation. Isolate the affected environment and stop promoting artifacts it produced. Use existing records and approved forensic tooling to investigate. Starting Python or package-management tools in that environment can activate the code.
- Found in a cache or mirror. Restrict reuse, retain the relevant evidence and trace which builds pulled the copy. A cached file is a lead. An installation or execution record gives you a stronger basis for scoping credential exposure.
- Installed version unclear. Focus investigation on jobs with privileged identities or production access. Use retained build metadata and identity-provider records to narrow the timeline before declaring a runner clean.
Revoke access while the rebuild is under way
From a trusted workstation, assign someone to each identity that the affected process could reach. Include cloud principals, package publishers, deployment accounts and human sessions where the evidence supports it. Revoke exposed keys at the service that accepts them. For sessions and federated identities, consult the provider's incident controls for active sessions, further token issuance and access that may have been granted during the window. Review sign-ins, role changes and artifact publications for actions outside the affected job's expected work.
Give replacement credentials only to rebuilt environments
- Build from a trusted base. Replace affected images, virtual environments and persistent runners. Rebuild with a reviewed lockfile and known-good artifacts from an approved source. Verify artifact digests against the approved release record.
- Keep old artifacts out of the new build. After preserving evidence, replace affected cache entries and quarantine packages your internal mirror still holds. Confirm the new runner resolves the intended package version and artifact instead of restoring an old cached environment.
- Restore the minimum required access. Run a validation job before reconnecting production credentials. Give dependency resolution only the access it requires, and make deployment identities available only to the jobs that deploy. Confirm provider-side revocation before closing the incident.
Handle human and machine handoffs separately
Put runtime values in the approved secrets manager for the rebuilt workload. If a verified responder needs a temporary text secret, a Secretus one-time share can pass it between trusted devices without placing the value in the incident channel. Record the recipient, scope and expiry in the incident log and keep the secret itself out of it.
