Update — November 19, 2026: New technical findings from Elastic
Elastic Security Labs has published additional analysis of the malicious agent-framework 1.26.1 and 1.26.2 releases. Its examination confirms that Python .pth startup hooks execute the credential collector whenever the interpreter starts in an affected environment—importing Agent Framework is not required. In version 1.26.1, the hook was agent_framework_meta.pth; in 1.26.2 it was renamed af_runtime_diag.pth, with a revised module and added GitHub Actions token targeting.
The reported collector read environment variables from as many as 300 processes and sought Azure CLI, pip, npm, Docker, Git and Hugging Face secrets. It also queried Azure managed-identity, Azure DevOps and GitHub Actions OIDC mechanisms to obtain cloud tokens accessible to the host, then attempted to send compressed, encoded data to a webhook.site endpoint. This broadens the necessary review from package versions alone to the identities and secrets available to any affected developer laptop or build agent.
Security teams should look for the two startup-hook filenames, determine whether Python launched while affected wheels were present, and inspect network and security telemetry using the Elastic findings and indicators. Pin to the previously clean version 1.26.0 while following current project guidance; rebuild affected environments, rotate exposed long-lived credentials and audit use of short-lived tokens that may not be directly revocable. The number of successful victim installations and downstream compromises has not been established publicly. This update is based on published malware analysis, not hands-on testing by IPCRIME.
Update — November 18, 2026
Microsoft repository member Vincent Biret confirmed in a November 17 comment on the official project issue that agent-framework 1.26.1 and 1.26.2 had been pulled from the package feed and the team was investigating. This acknowledgement confirms removal; it does not establish a completed root-cause investigation.
A GitHub Advisory Database entry published November 18 now classifies both versions as malicious and identifies 1.26.0 and earlier as unaffected by this package incident. The advisory lists no patched version: 1.26.0 is an earlier unaffected release, not a new remediation build. Removing the package or selecting an unaffected version does not resolve credentials already exposed by an executed payload.
Security researchers at Wiz and StepSecurity report that versions 1.26.1 and 1.26.2 of Microsoft’s agent-framework package on PyPI contained credential-stealing code. The releases appeared on November 17, 2026; StepSecurity published its analysis on November 18.
Two releases, then removal
Wiz places the first malicious upload at 20:35 UTC and the second at 21:08 UTC on November 17. It reports that both versions were unavailable on PyPI by about 21:30 UTC and identifies 1.26.0 as the last clean release in its investigation. Package removal does not remove copies already installed or retained in caches.
The researchers describe Python startup hooks that execute the payload when an interpreter starts in an affected environment, even without importing the framework. The code targets credentials available to the process, including cloud and development-pipeline access.
Review workload identities as well as stored secrets
StepSecurity’s analysis describes collection of environment variables and credential files, plus attempts to obtain tokens from workload identity services. That means a review limited to passwords saved in a repository could miss relevant access. Its November 18 report also examines the publishing-account activity, but a complete official root-cause account was not established by the sources reviewed for this article.
What affected teams should check
Wiz recommends checking lockfiles, CI logs, environments and images for the two affected versions, reviewing evidence of execution and rotating exposed credentials. StepSecurity likewise advises treating accessible secrets as exposed on affected systems. Use the researchers’ current indicators to guide a scoped investigation, including package caches and mirrors.
IPCRIME analysis: Keep prevention and recovery as separate tasks. Pinning a known clean dependency helps stop a repeat installation; rebuilding an affected environment and reviewing the identities it could use addresses earlier exposure. Record which jobs ran, what access they held and which destination systems require audit.
IPCRIME’s vendor-key and exposed-secret review guide provides related inventory questions. A local secret scanner cannot establish whether a stolen credential has already been used elsewhere.
Featured image: conceptual IPCRIME artwork, not a photograph or screenshot of this incident or product.
