AI Agent-Assisted Hackers Poisoned Microsoft's PyPI Package: 55 Minutes Was Enough to Steal Keys
Two poisoned versions of Microsoft's AI agent package were published to PyPI, the Python library catalog. The package is downloaded about 41,000 times a week. For 55 minutes, the versions sent cloud keys to the attacker on every Python launch, even without importing the package. Stolen Azure tokens remain valid for up to 24 hours, so removing the package did not help. Pin version 1.26.0 and reissue secrets if you installed 1.26.1 or 1.26.2.
AI-processed from X: @whyNexqor; edited by Hamidun News
On November 17, two poisoned versions of agent-framework appeared on PyPI, the public catalog from which Python developers install libraries with pip install. agent-framework is Microsoft's official package for building AI agents, and it is downloaded about 41,000 times a week. The malicious versions sat in the catalog for only 55 minutes, but that was enough. When installed, they began sending cloud keys and developer tokens to the attacker. If you installed this package this week, your secrets may have leaked. Wiz researchers disclosed the attack, and the author @whyNexqor summarized it on social media.
Why 55 Minutes Was Enough
The poisoned versions 1.26.1 and 1.26.2 contained a file with a .pth extension. Python reads such files every time the interpreter starts, so the package does not need to be imported. Think of it as a key planted in your apartment that opens the door every time you walk into the house. A single pip install is enough, and the code runs on any Python launch on the machine: in a script, in tests, in a CI job. No clicks are required after installation.
According to Wiz, the following data was leaking out:
- Cloud access: Azure CLI, Azure managed identity tokens, AWS keys, and OIDC tokens from CI;
- Developer tokens: GitHub, OpenAI, npm, PyPI, and Hugging Face, plus git credentials and Docker config.
Azure tokens remain valid for up to 24 hours. The package was removed from the catalog after roughly an hour, but that does not stop access using tokens that were already stolen.
How the Attacker Got Into the Release Pipeline
The breach was not a typical password guess against a PyPI account. According to Wiz, the attacker apparently stole a client certificate and the private key of Microsoft's release system. The first attempt to use them from AWS failed because IP address filtering blocked it. The second came from Microsoft's own address space and succeeded. The source does not say how the attacker obtained the certificate.
The lesson for teams protecting their own releases: the IP filter stopped the outside attempt but did not help once the attacker had a foothold inside the network. A network perimeter alone is not enough.
An Agent That Gave Itself Away
The unusual detail is that AI tooling sat behind the attack. Wiz found the config of a coding agent on the attacker's machine. It is an open agent harness running on a model from the GLM family. The tooling calls itself Cairn and is positioned as an open framework for AI pentesting, meaning automated vulnerability testing of systems. Wiz does not say who ran it.
The agent was careless. It sent progress reports to the same webhook where the stolen data went. Researchers read these requests and found data from at least three real victims. Automation made the attack easier to run, but it also left traces.
Caveat: this is an account from one research team, relayed on social media. The scale of damage, the number of victims, and the attacker's identity have not been publicly confirmed. For details, see the Wiz report and Microsoft's notifications.
What to Do Today
- Pin agent-framework to version 1.26.0, the last clean release.
- If 1.26.1 or 1.26.2 was installed anywhere, treat as compromised every secret that host or CI runner could read, and reissue them.
- Review Azure sign-in logs for managed identities and federated credentials.
- In CI, pin exact versions of all AI packages and use lock files with hashes. Auto-updating to the newest release on the day it ships is exactly how poisoned builds get in.
- For long-lived tokens, set minimal permissions and a short lifetime. A stolen Azure token that lasts 24 hours gives an attacker a large window.
For product leads and founders, the takeaway is simple. Tools for building AI agents have become attack targets, and attackers are now using agents against them. Dependencies from the AI ecosystem get installed quickly and often without review, and developer machines hold many credentials. Ask your team: who checks which AI library versions actually make it into your builds?
- Nov 21, 2026Microsoft научила Claude Code помнить прошлые сессии: стоимость задач ниже на 23,6%
- Nov 21, 2026ChatGPT cited Ahrefs pages, but in 43% of answers it recommended other companies' conferences
- Nov 21, 2026LangSmith launches Tool Gateway: agents reach tools through a single URL
- Nov 21, 2026Microsoft спрятала команду в ИИ-агенте, поменяв 7 нейронов: она сработала в 98,8% сессий
Key Questions Answered
Why 55 Minutes Was Enough?
The poisoned versions 1.26.1 and 1.26.2 contained a file with a .pth extension. Python reads such files every time the interpreter starts, so the package does not need to be imported. Think of it as a key planted in your apartment that opens the door every time you walk into the house. A single pip install is enough, and the code runs on any Python launch on the machine: in a script, in tests, in a CI job. No clicks are required after installation.
How the Attacker Got Into the Release Pipeline?
The breach was not a typical password guess against a PyPI account. According to Wiz, the attacker apparently stole a client certificate and the private key of Microsoft's release system. The first attempt to use them from AWS failed because IP address filtering blocked it. The second came from Microsoft's own address space and succeeded. The source does not say how the attacker obtained the certificate.
What are the key details on An Agent That Gave Itself Away?
The unusual detail is that AI tooling sat behind the attack. Wiz found the config of a coding agent on the attacker's machine. It is an open agent harness running on a model from the GLM family. The tooling calls itself Cairn and is positioned as an open framework for AI pentesting, meaning automated vulnerability testing of systems. Wiz does not say who ran it. The agent was careless. It sent progress reports to the same webhook where the stolen data went.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.