Technology·News & analysis
Hackers slipped a password stealer into Microsoft Agent Framework
Two fake releases of Microsoft's agent-framework package were live on PyPI for about an hour on November 17. They stole cloud passwords every time Python started.

Tide
Wave
Sci-fi
4/10
Reality
Shipping
A fake update that quietly raids the digital pantry.How we rate
Attackers published two malicious versions of Microsoft's agent-framework package to PyPI on November 17, 2026, each carrying a stealer for cloud passwords and tokens.
The code ran every time Python started, even when the package was never imported. The package gets about 192,000 downloads a month. PyPI removed both versions within about an hour, but Microsoft has not said how the attacker got in.
What to know
- Versions 1.26.1 and 1.26.2 of Microsoft's agent-framework were published to PyPI on November 17, 2026, 33 minutes apart, and both carried a credential stealer.
- The code ran every time Python started, even if the package was never used, and went after Azure, GitHub, AWS, OpenAI and other secrets.
- PyPI pulled both versions within about an hour, and the last clean release is 1.26.0. Microsoft has not yet said how the attacker got in.
- Anyone who installed either version should rotate every secret the machine could reach.
Attackers published two booby-trapped versions of Microsoft's agent-framework package on November 17, and both were built to steal passwords. So how does a trusted Microsoft download turn into a thief, and who needs to worry?
First, what is this package?
Microsoft Agent Framework is a free toolkit for building AI agents, programs that can plan steps and use tools on their own. It also helps several agents work together in one workflow.
The code is written in Python, and developers get it from PyPI. PyPI is the Python Package Index, a giant public shelf of ready-made code. A developer types one command, and the package lands on their machine.
The outer package, called agent-framework, pulls in the real framework code. The package gets about 192,000 downloads a month, according to the project's PyPI page as read by SafeDep. Wiz puts it at about 41,000 a week.
In real life Think of PyPI as an app store for programmers. If someone swaps a trusted app for a fake one with the same name, nobody notices until it has already run.
What happened on November 17?
Two bad versions went up that day, both on the official Microsoft project. Here is the timeline, in UTC:
- November 12, 18:25: the clean version 1.26.0 is published.
- November 17, 20:35: the first malicious version, 1.26.1, goes live.
- About 21:08: version 1.26.2 follows, 33 minutes later.
- About 21:30: both versions are no longer available on PyPI.
The second release is the worrying detail. PyPI had already pulled the first one, yet the attacker could still publish again. That means the attacker held real publishing access, not just a lucky one-off upload.
Several security companies spotted the bad versions quickly. Chainguard says its tooling blocked 1.26.1 within 48 seconds of publication. SafeDep flagged the second release 4 minutes after upload.
How did the stealer run without being used?
Normally, code only runs when you import it. These versions skipped that step.
Python has a small feature: when it starts, it runs any file ending in .pth that sits in its library folder. The attackers dropped one of those files in. It held a single line that launched the stealer.
That means installing the package was enough: the stealer ran every time Python started in that environment, even if nobody ever used agent-framework.
That includes quiet jobs such as running tests, using a code editor's helper tools or building software on a server.
The catch: the files were named to look like normal telemetry, the little reports software sends about its own health. The code even called itself "Lightweight runtime diagnostics hook", so a quick look would not raise alarms.
The second version used new file names. Anyone who searched only for the first version's file names would have missed it.
What did it steal?
The stealer was built for people who build AI and cloud software. It collected:
- Secret names in the environment: any setting whose name contains words such as TOKEN, SECRET, PASSWORD, KEY, AZURE, AWS, GITHUB or OPENAI.
- Other programs' secrets: it read the settings of up to 300 other running processes, where the system allowed it.
- Config files: pip, npm, Git and Docker login files, plus .env files where developers often keep keys.
- Azure identity tokens: a token is a short-lived digital pass that proves who you are, and the code asked Azure for these passes directly.
- GitHub Actions passes: the second version also asked for tokens from GitHub Actions, the service many teams use to build and test code automatically.
Everything was packed into one file and sent to a public webhook address, webhook.site. A webhook is just a web address that accepts data, and anyone can create one for free. Chainguard says the target was developer machines, build servers and live cloud workloads.
How did the attacker get in?
Microsoft has not said. Researchers have pieces of the answer, and they are not final.
Wiz says the attacker appears to have stolen a client certificate and private key for Microsoft's ESRP. That is the internal release service Microsoft uses to sign and publish this package. Wiz says the first publish attempt came from Amazon's cloud and was rejected by an IP filter. A second attempt, from Microsoft-owned addresses, worked. How the credentials were obtained is not known.
SafeDep found another clue. A PyPI account named vibiret, created the same day as the attack, is listed as a maintainer of agent-framework and five related projects. SafeDep says the join date points to that account as the uploader, but PyPI and Microsoft have not published the upload log that would confirm it.
Why it matters: if that account kept its access, a new bad release could appear at any time. Four of those five related projects install together with agent-framework.
Is this a pattern?
It is the second Microsoft Python package to get rogue releases in about six months. In May, three versions of a package called durabletask were published after its PyPI token was stolen. SafeDep and Wiz both say the payloads and servers differ, so there is no evidence that the two cases are linked.
Wiz also found something odd in the attacker's own tooling. It reported its progress to the same open webhook it used for stealing, which let Wiz watch part of the operation. The tooling named itself Cairn, an open-source AI penetration-testing framework.
SafeDep also found that earlier releases came with two files, a wheel and a source archive, while the bad ones were a single wheel. Missing proof of origin did not help either: PyPI has no signed record for 1.26.0, so users had no cryptographic way to tell Microsoft's uploads from fake ones.
What should you do now?
Developers and security teams can check quickly. Chainguard and BleepingComputer's advice lines up:
- Search: look in lockfiles, caches and build logs for agent-framework 1.26.1 or 1.26.2.
- Isolate: if a machine installed either version, avoid starting Python on it until it is cleaned.
- Rebuild: go back to a clean environment with version 1.26.0.
- Rotate: change every credential the machine could reach, and revoke GitHub Actions and Azure DevOps access.
- Check logs: look for connections to webhook.site, and tell blocked ones from completed ones.
Pin versions where you can. Pinning means telling your tools to install one exact version, so a surprise new release is not pulled in automatically. Treat any new agent-framework release from November 17 or 18 with suspicion until Microsoft confirms the package is safe.
What it means for you
- If you build with agent-framework: check your installs today, even if the build finished days ago.
- If you only use apps: nothing to do, unless a developer you rely on shares the news about their own tools.
- If you run a team: keep long-lived keys out of build machines, since a stealer can only take what is within reach.
- If you follow AI tools: the new toolkits for agents are becoming a target, because they sit next to powerful keys.
The bottom line
A trusted Microsoft package was hijacked for about an hour, and that was long enough to matter. The fix for most people is simple: check for 1.26.1 and 1.26.2, then rotate secrets if you find them.
The open question is how the attacker got in, and Microsoft has not answered it yet. Until it does, treat new releases with care.
Key facts
- Bad versions
- 1.26.1 and 1.26.2
- Published
- November 17, 2026
- Last clean version
- 1.26.0
- Package downloads
- About 192,000 a month
- Root cause
- Not yet confirmed by Microsoft
Got questions?
Quick answers, plain wordsWhat happened to Microsoft's agent-framework package?
Two malicious versions, 1.26.1 and 1.26.2, were published to PyPI on November 17, 2026. Both carried a stealer that collects passwords, tokens and keys and sends them to the attacker.
Which versions are safe?
Version 1.26.0 is the last clean release. Anything published after it on November 17 or 18 should be treated with suspicion until Microsoft says the package is secure again.
What is PyPI?
PyPI is the Python Package Index, the public store where Python developers download ready-made code libraries. It is the default source when you run the pip install command.
Do I need to import the package for the stealer to run?
No. The malicious code sat in a startup file that Python runs every time it launches in that environment, so installing the package was enough.
What did the malware steal?
It collected environment variables with names such as TOKEN, SECRET, PASSWORD and KEY, plus config files for pip, npm, Git, Docker and Azure. It also requested fresh Azure and GitHub Actions identity tokens.
How do I know if I was affected?
Search lockfiles, caches and build logs for agent-framework 1.26.1 or 1.26.2. Security teams also suggest looking for connections to webhook.site in network logs.
What should I do if I installed a bad version?
Isolate the machine, rebuild from a clean environment using 1.26.0, and rotate every credential that machine could reach. Revoke any GitHub Actions and Azure DevOps access it held.
Did Microsoft say how the attacker got in?
Not yet. Wiz says the attacker appears to have used a stolen certificate and key for Microsoft's internal release service, but how they got them is unknown.
Is this the first time a Microsoft package was hit?
No. A Microsoft package called durabletask got rogue releases in May after an API token was stolen. Researchers have not linked the two incidents.
Is the Microsoft Agent Framework itself unsafe now?
The framework's code on GitHub is not reported as changed, and the core package stayed at the clean 1.26.0. The problem was the two bad releases of the outer package.
SourcesWiz Research
Topics and tagsMicrosoft, AI agents, Cybersecurity, microsoft
Go deeper
Related guide
Guide, 45 minHow to build a daily AI agent with Claude Managed AgentsAnthropic's open example is an agent that reads Slack and GitHub each weekday and posts one short brief. Here is how it works and how to...Related stories

SignDrop, a new Windows zero-day, gives attackers full System rights
A researcher with a long feud against Microsoft published SignDrop, an unpatched Windows flaw that turns a normal account into the most powerful one on the PC.

Chainalysis says hackers took logins and tokens for all its customers
The crypto-tracing firm says an attacker used AI agents, then an on-path attack, to take names, password hashes and live tokens for every user. Its tools are back online.

Microsoft takes down a phishing marketplace behind 1 billion scam messages
Microsoft and Health-ISAC seized domains behind YAM Store, CoGUI and Iron Man, three services that sold scammers accounts, fake websites and ready-made lures used in more than 1 billion messages this year.
More in brief
- ASELSAN's EJDERHA electromagnetic weapon disables 51 drones with one commandNov 21
- Microsoft To Do will shut down on June 30, 2027Nov 21
- Medtronic's Hugo surgical robot cleared for general surgery in the USNov 21
- Google's Gemini now turns noisy fire radio into text for Watch DutyNov 21
- Internet Archive staff move to unionize with CWANov 20
- Etsy sellers are leaving over new terms that allow AI trainingNov 20