Threat

Microsoft's official Python agent framework hijacked to distribute malware

Page published · Page updated

Dossier

Earliest dated coverage: 18 Nov 2026 · First observed: 18 Nov 2026 · Latest dated coverage: 18 Nov 2026

Coverage timeline

18 Nov 2026theregister.com

Single-source incident — one report is available.

Why it matters

Compromise of Microsoft's official AI-agent framework turns a trusted dependency into a credential-theft vector for developer workstations and CI environments, making the reported versions an immediate investigation priority.

Microsoft's official agent-framework PyPI package was compromised to distribute credential-stealing malware in versions 1.26.1 and 1.26.2, according to the report. The malicious package executes when Python starts without requiring an import, exfiltrates environment variables and developer secrets, and installs the genuine 1.26.0 framework to maintain legitimate functionality. Targeted credentials include those for Azure, Git, Docker, and Hugging Face.

exploited-vuln

Summary

Microsoft's official agent-framework package on PyPI was reportedly hijacked to distribute credential-stealing malware in versions 1.26.1 and 1.26.2. The Register attributes the findings to OX Security and OpenSourceMalware. All substantive incident details in the supplied evidence come through this single news report; the underlying research documents are not supplied for independent corroboration.[0]

Exposure does not require an application to import agent-framework. The malicious package reportedly runs when a Python interpreter starts in an environment where the package is installed, while retrieving the genuine 1.26.0 release to preserve legitimate functionality. It collects environment variables and credential files, actively requests fresh cloud and CI identity tokens, and sends collected information to a hard-coded webhook.site endpoint.[0]

The reported release sequence raises a containment concern: OX Security observed version 1.26.1 reverted to a safe version, followed by an attacker-uploaded 1.26.2 release through Microsoft's official PyPI publisher account. This indicates that replacing one malicious release had not, at that point, prevented further unauthorized publication. The report suggests possible GitHub organization administrator compromise, but does not establish how the attacker initially obtained access.[0]

Defenders should prioritize identifying affected environments, removing malicious package artifacts, and responding to potential exposure of developer, cloud, and CI credentials. Package removal addresses the reported execution mechanism but does not undo credential theft or fresh-token issuance. As of the article's publication, final remediation was not established, and the separately reported ONNX Runtime repository outage had not been linked to this incident.[0]

Attack chain

  1. Malicious release distribution: The attacker distributed compromised agent-framework releases through the official package channel. The reported sequence was a malicious 1.26.1 release, its reversion to a safe version, and a subsequent malicious 1.26.2 upload through Microsoft's official PyPI publisher account. The initial means of obtaining publishing access is not established.[0]
  2. Installation with legitimate functionality retained: The malicious package retrieves the genuine agent-framework 1.26.0 release, allowing expected framework functionality to remain available while the credential-stealing behavior operates in the background.[0]
  3. Execution at Python startup: Once the malicious package is installed, starting a Python interpreter triggers its behavior without requiring an explicit agent-framework import. The report ties continued operation to the malicious package's presence in a site-packages directory.[0]
  4. Credential collection and token acquisition: The malware gathers environment variables, credential files, and service secrets, including material associated with Azure, Git, Docker, and Hugging Face. It also requests fresh identity tokens from GitHub Actions, Azure DevOps, and Azure. The report does not establish an internal ordering between these collection methods.[0]
  5. Exfiltration: Collected information is packaged as base64-encoded JSON and sent to a hard-coded webhook.site endpoint. The same destination reportedly receives both existing credentials and newly requested identity tokens.[0]

Disclosure timeline

DateEvent
2026-11-17The Register's November 18 account places OX Security's reporting on the compromised releases late on the preceding day. It attributes the 1.26.1 reversion and subsequent malicious 1.26.2 upload sequence to OX Security, without giving precise times for those release events.[0]
2026-11-18The Register describes OpenSourceMalware's reporting as published that morning, including analysis of existing-secret theft and active identity-token requests.[0]
2026-11-18T17:23:00ZThe Register published its report. The agent-framework and ONNX Runtime GitHub repositories were reported offline, no connection between the ONNX Runtime outage and the malware activity had been established, and Microsoft had not responded to the publication's questions.[0]

How it works

The documented mechanism is malicious modification of an official package distribution, rather than exploitation of a described programming flaw. The significant access observation is the reported upload of malicious version 1.26.2 through Microsoft's official PyPI publisher account after a preceding release was reverted. The suggestion of compromised GitHub organization administrator credentials remains an unconfirmed explanation for initial access.[0]

Execution is tied to Python interpreter startup rather than application-level use of the framework. Consequently, an environment containing the malicious release can be exposed even if its applications never import agent-framework. The supplied article does not identify the specific startup-hook implementation or malicious filenames, so no particular hook or file can be asserted.[0]

Credential access combines passive collection with active acquisition. The malware reads available environment variables and credential files, then also requests new identity tokens from GitHub Actions, Azure DevOps, and Azure. The cited researchers emphasize that immediate exfiltration can make even short-lived tokens useful to an attacker in cloud build environments.[0]

The report describes exfiltration of base64-encoded JSON to a hard-coded webhook.site endpoint. Although the article also uses the term 'encrypts,' it supplies no separate cryptographic details sufficient to characterize an encryption mechanism. It reports no separate persistence mechanism beyond the malicious package remaining installed; that limitation does not establish that previously stolen credentials are no longer usable.[0]

Affected versions and patch status

ProductAffectedPatch status
Microsoft agent-framework for Python, distributed through PyPIVersions 1.26.1 and 1.26.2 are identified as compromised releases. The report also says 1.26.1 was subsequently reverted to a safe version, but supplies no artifact hashes or precise replacement times to distinguish release contents.Version 1.26.0 is described as the genuine release retrieved by the malicious package. Reversion of 1.26.1 was followed by a malicious 1.26.2 upload. The supplied report does not establish final publisher-account containment, a definitive fixed release, or completed remediation.[0]

Indicators of Compromise

TypeIndicatorContext
domainwebhook.siteThe reported host for a hard-coded credential-exfiltration endpoint. No unique endpoint identifier is supplied. This is a shared service domain, so a domain match alone is insufficient to attribute traffic to this incident; correlate with affected package versions and Python execution.[0]
otheragent-frameworkPackage-inventory hunting identifier. Investigate installations or installation history involving versions 1.26.1 and 1.26.2. The package name alone is not malicious, and the reported reversion of 1.26.1 means version labels alone may not distinguish malicious and replacement artifacts.[0]

Key takeaways

  • Import-based exposure checks are insufficient for this incident: the reported execution trigger is Python interpreter startup in an environment containing the malicious package.[0]
  • The credential-response scope extends beyond stored secrets. Fresh identity-token requests make CI and cloud identity activity central to investigation and containment.[0]
  • Release rollback and publisher-access containment are distinct problems. The reported progression from a reverted 1.26.1 release to malicious 1.26.2 demonstrates why both must be addressed.[0]
  • Attribution and impact remain bounded: the supplied evidence does not establish the initial-access mechanism, a named attacker, downstream use of stolen credentials, or a connection to the ONNX Runtime repository outage.[0]

Defensive actions

  • Identify environments that installed agent-framework 1.26.1 or 1.26.2, including developer environments and CI runners; do not limit scoping to applications that imported the framework.: Both versions are reported as compromised, and malicious execution occurs at Python startup without an agent-framework import. The article explicitly advises treating either version's presence as requiring account-security response.[0]
  • Remove malicious package artifacts from affected environments and validate the provenance of any replacement rather than relying only on its version label.: The reported malware operates while the malicious package remains in site-packages. Because 1.26.1 was reportedly reverted and then followed by another malicious release, neither a familiar version label nor a single release rollback establishes complete containment.[0]
  • Revoke or rotate potentially exposed credentials available to affected hosts and runners, covering environment variables and credential files for Azure, Git, Docker, Hugging Face, and other accessible developer services.: The malware reportedly steals these classes of secrets. Removing the package does not retract information already transmitted to the attacker.[0]
  • Investigate token issuance and potential misuse in GitHub Actions, Azure DevOps, and Azure, and invalidate exposed tokens where supported.: The malware does not rely solely on credentials already present: it actively requests fresh identity tokens and exfiltrates them immediately. The report specifically identifies this behavior as dangerous in cloud build environments.[0]
  • Review outbound telemetry from affected environments for connections to webhook.site, correlating matches with malicious-package presence and Python startup activity.: The report identifies a hard-coded endpoint on this host as the exfiltration destination. Because the supplied evidence lacks a unique endpoint identifier, correlated investigation is more reliable than treating every connection to the shared domain as malicious.[0]
  • For maintainers responding to the publishing compromise, secure publisher access and verify replacement artifacts before treating a release rollback as containment.: The reported upload of malicious 1.26.2 through the official publisher account after 1.26.1 was reverted shows that replacing a malicious release alone had not prevented further unauthorized publication at that stage.[0]