Two poisoned releases of Microsoft’s Agent Framework Python package - what happened?

Resumen detallado
Two malicious releases reached the legitimate Python package listing. Their removal stops fresh downloads from that source, but affected development and build environments...
Andrew Matler, Editor in Chief and CEO | Published 19 November 2026, 3:45 pm AEDT
Two malicious releases reached the legitimate Python package listing. Their removal stops fresh downloads from that source, but affected development and build environments still need investigation.
Two poisoned releases of Microsoft’s Agent Framework Python package, software for building AI agents, could run credential-stealing code before an application imported the framework, making ordinary interpreter startup a potential exposure event for developers and automated build systems.
The GitHub malware advisory identifies `agent-framework` versions 1.26.1 and 1.26.2 as malicious. This was the legitimate package listing, not a similarly named imitation. The advisory identifies 1.26.0 as the last known clean release.
According to Wiz’s investigation, the releases appeared on 17 November at 20:35 UTC and 21:08 UTC respectively, or 7:35am and 8:08am AEDT on 18 November. Both were unavailable on PyPI by 21:30 UTC. The current project listing shows 1.26.0.
The short distribution window is not the same as a short exposure window. A downloaded wheel, cached build dependency or container image can outlive removal from the public index. Organisations need to establish what their systems actually installed and ran, rather than infer safety from today’s package listing.
The startup mechanism explains why that distinction matters. The malicious distributions placed `.pth` files in Python’s package directory. These files can contain executable import lines, processed during normal startup by Python’s `site` machinery. Python’s documentation describes this behaviour. An application did not need to call an Agent Framework function for the payload to run in an affected environment.
This is a software delivery compromise, not evidence that an AI model went rogue. Reviewing an agent’s prompts or tool permissions would not, by itself, identify a malicious interpreter startup hook.
Elastic’s analysis describes…
Inicio del artículo, extracto de iTWire. Para leer el resto, abra el artículo en el sitio del medio.
MarketNews muestra el titular y el resumen publicados por el medio. El artículo completo se lee en el sitio del editor. (itwire.com)


