Digital Transformation · 19 November 2026
Microsoft Agent Framework PyPI package hijacked to steal Azure and CI credentials
Attackers briefly took over Microsoft's Agent Framework package on PyPI, publishing tampered versions 1.26.1 and 1.26.2 on 17 November that harvested cloud tokens and CI/CD credentials for less than an hour.
What happened
Attackers briefly took over Microsoft's official Agent Framework package on the Python Package Index (PyPI) and used it to distribute credential-stealing malware, according to several security researchers. Two tampered releases, versions 1.26.1 and 1.26.2, appeared on 17 November and remained available for less than an hour.
The malicious code was built to collect cloud tokens, developer secrets and CI/CD credentials from developer workstations and automated build systems, including Azure managed identity tokens. It then sent them to an attacker-controlled webhook. OX Security, one of the firms that flagged the releases, reported that the payload ran every time Python started, even when no code imported the package.
The tampered versions also installed the genuine framework code, so applications continued to work as expected. According to the InfoWorld reporting, that design is what made the compromise easy to miss.
Why it matters
Agent frameworks are becoming a core layer in how organisations build AI-driven workflows, which means they can sit close to the most privileged parts of the software supply chain. An orchestration package can touch model endpoints, cloud identities and build pipelines. When that layer is compromised, the damage is not confined to one application. It can reach the credentials that let teams deploy and run everything else.
For digital transformation leaders, the incident shows that AI adoption widens the attack surface as quickly as it expands capability. Teams moving fast on agent tooling may be extending trust to dependencies they have not reviewed as rigorously as core infrastructure. Because the malicious build still functioned, standard functional testing would not have flagged it, which points to the need for provenance and behavioural checks rather than relying on whether software appears to work.
By the numbers
- 2 tampered versions, 1.26.1 and 1.26.2, were published to PyPI.
- 17 November is the date the malicious releases appeared.
- Under one hour is how long the packages remained available.
- 192,534 is the usage figure cited for the package in the reporting.
The Renascence take
Most coverage will treat this as a single bad upload with a short window. The more useful reading is structural. The trust developers place in package registries is now part of the experience of any AI programme, because engineers are the first users of every agent an organisation ships.
The lesson is not that one upload slipped through. It is that a dependency adopted for speed has quietly become a privileged actor in your environment. Treat agent frameworks the way you would treat a payment gateway: pin versions, verify provenance and limit what a build job can reach. Behavioural economics would call this a small, upfront friction that avoids a far more expensive surprise later, and operators who wait for an incident to justify it will pay for it on a worse timeline.
Sources
This briefing was written by our Newsdesk, synthesising reporting from the outlets below. Follow the links for the original coverage.
FAQ
Questions we get on this topic
More in Digital Transformation
Stay ahead of CX
Get the signal, not the noise.
The stories shaping customer experience — plus the Journal and Experience Loom — in your inbox.


