Did attackers actually publish malicious Microsoft Agent Framework packages on PyPI?
Yes. Security researchers documented malicious versions 1.26.1 and 1.26.2 of Microsoft’s agent-framework Python package, released on November 17, 2026. Both used startup hooks to attempt credential theft and were removed or quarantined; the full downstream victim count remains unverified.
- Malicious versions 1.26.1 and 1.26.2 appeared on PyPI at approximately 20:35 and 21:08 UTC on November 17; researchers identify 1.26.0 as the last known clean release.
- The injected .pth files could launch a credential stealer at Python interpreter startup without a direct import, targeting Azure, CI and other developer secrets.
- Wiz researchers examined 50 requests captured at the attacker’s public webhook and identified three unique successful victim-data exfiltration cases in that visible sample. This is not a count of every potentially exposed installation.
The signal
A real compromise of an AI-agent development dependency has turned software supply-chain security into a direct accountability problem for the agent ecosystem. On November 17, 2026, unauthorized releases of Microsoft’s agent-framework package appeared under its legitimate name on the Python Package Index (PyPI). Their versions were 1.26.1 and 1.26.2.
According to the GitHub Advisory Database and original research by Wiz, the altered packages contained credential-stealing code. The releases were removed or quarantined shortly afterward. The last known clean version at the time was 1.26.0. A package being genuine in name is not, by itself, evidence that every uploaded build is authorized or safe.
This was a documented malicious package publication, not merely a hypothetical prompt-injection scenario. But the event does not establish that Microsoft's main source repository or every deployment of its agent framework was compromised.
How the compromise worked
The malicious package placed an executable .pth file into Python's site-packages directory. Python processes these startup files when an interpreter launches, so the stealer could run in an affected environment even when application code did not explicitly import agent_framework. This makes ordinary dependency installation and subsequent Python execution a sensitive boundary.
Researchers describe attempts to collect environment secrets, Azure identity material, CI/CD credentials and other developer access tokens. The later 1.26.2 variant expanded coverage to certain GitHub Actions and OIDC contexts. The payload attempted to send captured information to an external collection endpoint.
Wiz's investigation indicates that attackers may have gained access to an ESRP publishing client certificate and key and used Microsoft-associated infrastructure to publish the packages. The initial credential-acquisition route remained unknown in its report. Cortist treats that explanation as a research finding and hypothesis about the publishing chain, not a complete publicly confirmed root-cause investigation.
Wiz examined 50 stored requests at the observed webhook endpoint, of which 17 represented traffic associated with three distinct successful victim-data exfiltration cases in the visible sample. Some other requests were linked to the attacker or scanning environments. Those figures should not be confused with total download counts, total victims or all actual credential exposures.
What can and cannot be concluded
The GitHub advisory lists 1.26.1 and 1.26.2 as malicious and does not designate a patched version. Instead, 1.26.0 or earlier are described as not affected by this package poisoning. Removing the uploads reduces new exposure but does not undo secrets that may already have left previously affected environments.
For a user who installed an affected wheel, the relevant follow-up is exposure review, rebuilding from a trusted base and rotation of reachable credentials and cloud or CI trust relationships. This is a security incident where outcome evidence must be granular: malicious upload, package installation, payload execution, observed exfiltration and verified downstream abuse are distinct events.
A compressed timeline and a legitimate project name make the case important for the governance of AI agents. Yet the package attack did not require the agent itself to behave autonomously or escape model alignment controls. It exploited distribution and runtime infrastructure.
04 / What Cortist will watch
A confirmed Microsoft or package-index incident postmortem; evidence of how publishing access was obtained; any additional confirmed compromised dependencies; the number of affected installations and validated exfiltration incidents; and independent verification that publishing controls and monitoring have been strengthened.
Future updates should distinguish completed remediation from merely removing the malicious packages. An official root-cause report, dated containment actions and independent tests would materially strengthen the record.
Frequently asked questions
Were all Microsoft Agent Framework installations infected?
No. The identified malicious releases were versions 1.26.1 and 1.26.2. The public evidence does not establish that every installation of the project received those versions.
Could the stealer run without an explicit package import?
Yes. Its .pth startup hook could execute when a Python interpreter started in an environment containing an affected package.
How many victims were confirmed?
Wiz identified three unique successful victim-data exfiltration cases within the subset of webhook requests it analyzed. The total affected population remains unverified.
Primary sources and verification limits
Cortist distinguishes documented software compromises and policy recommendations from independent proof of downstream damage, enacted duties or completed mitigation. The original Ledger baselines and official forecast scores are not modified by this analysis.