Attackers put credential stealers in Microsoft's AI agent package, and an hour was enough

4 Sources

Share

Attackers published two malicious versions of Microsoft's official agent-framework package on the PyPI repository on November 17, 2026, turning a package trusted by AI developers into a credential stealer

3

. Because it sat under Microsoft's own name, anyone or any build system that installed it during the window it was live could have been exposed

3

.

Source: Hacker News

Source: Hacker News

Agent Framework is Microsoft's open-source toolkit for developing and orchestrating AI agents in Python

2

. It draws roughly 190,000 to 200,000 downloads a month, per varying counts from the security firms

2

. That download volume means the pool of potential victims is large, though the firms documenting the attack have not confirmed any actual infections

2

. The two versions were live for roughly 45 minutes in total before PyPI took them down, per Aikido Security's timestamps

3

.

What the stealer grabbed

The malware was built to run whether or not a developer ever used the package. It hooked Python's startup process, so it executed every time Python started in an affected environment

3

. It also installed the genuine previous version of the package's code, so applications kept working normally and gave victims no obvious signal that anything was wrong

4

.

Once running, it harvested environment variables whose names matched words like token, secret, password, key, AWS, GitHub and OpenAI

2

. It read up to 60 credential files, from Azure CLI and Git configs to npm, pip and Hugging Face tokens, per OX Security

3

. It also pulled secrets from up to 300 running processes, per OX Security

3

. Then it went further and asked the cloud platforms to mint new tokens

1

. It requested Azure DevOps federated identity tokens and Azure managed-identity tokens for services including Key Vault and Microsoft Graph, plus, in the second version, GitHub Actions tokens

1

. "Besides reading credentials already on disk, the code asks the platform to issue new ones. Rotating stored secrets does not cover those," per SafeDep

2

.

Source: InfoWorld

Source: InfoWorld

Everything went to a single hard-coded webhook.site endpoint, disguised as a diagnostics module, which gives defenders a concrete string to search their logs for

4

. Endor Labs assessed that the payload was written mainly for automated build systems rather than developer laptops: "Developer laptops are in scope, but the payload is written for CI runners and Azure workloads"

3

.

The attacker came straight back

PyPI quarantined the first malicious version, and the attacker uploaded the second about two and a half minutes later with the payload hidden along a different file path

4

. Publishing access had survived the quarantine

4

. OX Security called it "a planned, live and coordinated effort to reach as many victims as they can, while still being able to bypass blocks and restrictions"

3

. At least seven security firms independently documented the compromise, including OX Security, Aikido, Endor Labs, SafeDep, Socket, StepSecurity and Google-owned Wiz

2

.

It is the second Microsoft-related PyPI package compromised this year, after three trojanized versions of the durabletask package appeared in May, though no researcher has linked the two attacks

2

.

Cleanup is more than an uninstall

The malware was not persistent: it ran only while the malicious version remained installed. Researchers at OpenSourceMalware noted that immediate exfiltration makes even short-lived tokens useful to an attacker

1

. And deleting the package recovers nothing. "Removing the package stops future runs, but it doesn't undo anything that was already sent," per Aikido Security

3

.

The firms' guidance reflects that: isolate affected hosts, pin to the verified 1.26.0 release with hash checking, and block the webhook.site endpoint

4

. They also advise rotating Azure, GitHub, npm, PyPI and other credentials, then auditing Azure sign-in logs along with GitHub Actions workflows

4

. Endor Labs warned that container images, continuous integration caches and lockfiles built during the window could still contain the malicious version, so cleaned-up repositories do not mean cleaned-up machines

3

.

OX Security believes attackers hijacked the wheel-file build process

3

; The Register reports it appears to involve a compromised administrator account on Microsoft's GitHub organization

1

. Neither theory is confirmed; without knowing the entry point, Microsoft and its users cannot be sure the attack path is closed

1

. Microsoft did not respond to press inquiries before publication

4

. The agent-framework GitHub repository was offline as of The Register's writing, which suggests Microsoft was actively responding

1

.

Source: BleepingComputer

Source: BleepingComputer

© 2026 TheOutpost.AI All rights reserved