Skip to content
Wednesday, November 18, 2026Free guide
EN
dailyai
Security

Wiz Finds Only 3 Victims in 55-Minute Microsoft AI Toolkit Hack

Wiz Research says only three unique victims had data stolen after two malicious versions of Microsoft’s AI agent toolkit sat on the main Python package index for about 55 minutes on 17 November.

· 2 min read
ShareXin

Wiz Research says only three unique victims had their data stolen after two malicious copies of Microsoft’s AI agent toolkit sat on PyPI, the main public library of Python software, for about 55 minutes on 17 November.

The cloud security firm published a first report on Tuesday and an update on Wednesday, after it inspected the 50 requests stored by the public inbox the attackers used. The package, called agent-framework, is Microsoft’s official Python toolkit for building AI agents. The official package page now shows 195,912 downloads in the past month.

The first malicious version, 1.26.1, went up at 20:35 UTC. A second, 1.26.2, followed 33 minutes later, at 21:08 UTC. By 21:30 UTC, both had been taken off the index. The last safe release is 1.26.0, published on 12 November.

What the fake updates did

Once either version was installed, hidden startup files ran whenever Python started on that machine, even if the toolkit itself was never used. The extra code copied cloud and build-system credentials, including Azure logins and, in the second version, GitHub Actions tokens, and sent them to a public webhook service.

The firm says the attacker appears to have used a stolen certificate and key for the internal Microsoft service that signs and publishes the package. A first upload, from Amazon’s cloud, was blocked by an address filter. A second, from a Microsoft-owned address, succeeded. How those signing credentials were obtained is not yet known.

Why the damage was limited

The inbox used to collect the stolen data was on the free tier of a public request-inspection service, which accepts at most 100 requests and expires after seven days. A maximum of 50 requests were stored. Four came from the attacker’s own machines, and 29 appeared to come from security sandboxes, vendor scanners and tests. The remaining 17 requests, the firm said, represent only three unique, successful cases of stolen victim data.

A separate technical analysis of both versions reached the same conclusions about what was collected and where it was sent. Earlier this week, a separate report described AI hacking agents that stole paid AI access keys and then used them to fund more attacks.

Anyone who installed 1.26.1 or 1.26.2 should pin the package to 1.26.0, rebuild the environment, and rotate credentials the machine could reach. Stolen cloud identity tokens can remain valid for up to 24 hours, the firm said.

Why it matters: A widely used Microsoft toolkit was hijacked through the company’s own signing pipeline, and the copies sat in the public library that Python developers trust. A short window and a free inbox cap limited the confirmed victims to three this time, but anyone who installed those versions still needs to replace their credentials.