Data from three victims reached the attacker behind the poisoned Microsoft Agent Framework releases, based on logs analyzed by Wiz Research. Versions 1.26.1 and 1.26.2 of the agent-framework package reached PyPI on November 17, 2026. Both carried a credential stealer aimed at developer machines, Azure workloads and CI pipelines.
The bad releases were live for less than an hour. Still, anyone who installed them should treat every reachable secret as exposed. The latest safe version is 1.26.0.
What happened on November 17
Microsoft Agent Framework is the company’s open-source toolkit for building AI agents and multi-agent workflows. The agent-framework package on PyPI is its Python meta-package.
According to the Wiz Research timeline, version 1.26.1 went up at 20:35 UTC. Version 1.26.2 followed about 33 minutes later, at roughly 21:08 UTC. Wiz says both were gone from PyPI by about 21:30 UTC.
Other researchers give slightly different end times. CSO Online cites Aikido Security, which says both versions were gone by about 21:20 UTC. BleepingComputer reports that the second release appeared about 2.5 minutes after PyPI quarantined the first. It adds that the second was quarantined about eight minutes later.
Both malicious builds still pulled in the genuine agent-framework-core 1.26.0. Apps kept working, so the compromise was easy to miss. Wiz found no matching release tag in Microsoft’s GitHub repository and no malicious code on the main branch.
Download figures vary by source. Wiz cites almost 200,000 downloads a month in one place and about 41,000 a week in another. OX Security puts the figure at 192,534 monthly downloads.
How the attacker got publishing access
Wiz updated its report on November 18 with a root-cause finding. The attacker appears to have compromised a client certificate and private key for Microsoft ESRP. ESRP, short for Engineering System Release Pipeline, is the internal service Microsoft uses to sign and publish this package.
The first publish attempt came from AWS. ESRP’s IP filtering rejected it. A second attempt from Microsoft-owned IP space succeeded. Wiz says it does not yet know how the credentials were obtained.
Wiz could watch part of the operation because the attacker’s tooling reported its progress to the same webhook used for stolen data. The tool calls itself Cairn, an open-source AI penetration-testing framework. It staged the package in /tmp/cairn_pkg and sent a probe named CAIRN_ESRP_PROBE2.
The attacker also tested the stealer on their own machine first. That run captured an AI coding agent setup that Wiz believes is the open-source “Pi” harness, backed by a GLM-family model.
Other theories exist but remain unconfirmed. The Register described what appeared to be unauthorized access to an administrator account in Microsoft’s GitHub organization. SecurityWeek reports that SafeDep flagged a PyPI maintainer account named “vibiret,” created on the day of the attack. Its role in the uploads is not confirmed.
What the attacker actually received
The stealer sent everything to one hard-coded webhook.site address. That choice limited what the attacker could collect. On the free tier, an unattached URL accepts at most 100 requests. After that, it returns an error and logs nothing.
Wiz confirmed the attacker used the free tier. It analyzed the 50 requests that the service stored:
- 4 came from the attacker’s own machines and attack process.
- 29 appear linked to sandboxes, vendor scanning and tests.
- 17 represent 3 unique, successful thefts of victim data.
Wiz has not named the victims.
What changes for you
The risk depends on where the bad version ran. A single install was enough, because the code ran at every Python start, even if your app never imported the package.
- Developers: Local tokens and config files were in scope, including Azure CLI sessions, git credentials, npm, pip and Docker settings, and Hugging Face tokens.
- Azure workloads: On VMs, App Service and similar hosts, the stealer asked Azure for fresh managed identity tokens for 12 resources. These include Graph, Key Vault, Storage and Container Registry.
- Azure DevOps pipelines: It read agent credential files and requested OIDC tokens when pipeline variables allowed it.
- GitHub Actions: Only 1.26.2 targeted Actions. In workflows with
id-token: write, it requested OIDC tokens for Azure.
Wiz notes that stolen managed identity and OIDC-exchanged tokens stay valid for up to 24 hours. Rotating stored secrets does not cancel tokens the stealer requested fresh, as SafeDep noted in comments cited by The Hacker News. Audit the identities behind them as well.
How to check whether you are affected
- Search for the versions. Look in lockfiles, CI logs, virtual environments and container images for
agent-framework1.26.1 or 1.26.2. Pin to 1.26.0. - Look for the hook files. In site-packages, check for
agent_framework_meta.pthandagent_framework_meta/_telemetry.py(1.26.1). Also check foraf_runtime_diag.pthandaf_runtime_diag.py(1.26.2). Searching only for the first set will miss the second. - Check outbound traffic. Search proxy and DNS logs for the webhook.site URL in Wiz’s report. A 200 response means the data was accepted. A 410 or 429 suggests the service rejected it.
- Check build caches. Endor Labs, quoted by CSO Online, advises checking images, CI caches and lockfiles built during the exposure window.
- Isolate before cleaning. BleepingComputer advises not running Python or pip on an affected system until it is secured, because any Python start runs the payload again.
What to rotate
Wiz recommends rotating every secret in the environment on affected hosts and runners. That covers Azure CLI sessions, Azure DevOps personal access tokens and service connections, and git credentials. It also includes npm, PyPI, NuGet, Docker and Hugging Face tokens.
Next, review Azure sign-in and activity logs for managed identities and federated credentials. For 1.26.2 in GitHub Actions, review Entra ID federated credentials that trust the affected repositories. Then rotate repository and organization secrets exposed to those jobs. OX Security also suggests blocking outbound traffic to webhook.site from build systems.
Removing the package stops new runs, but it does not undo past theft. Aikido Security made the same point in comments quoted by CSO Online.
Technical details: how the stealer worked
Python reads .pth files in site-packages every time the interpreter starts. A line in such a file that begins with import runs as code. Both malicious versions used this trick, so the stealer started without any action from the app.
The code tried to look like normal SDK telemetry. It used names such as _telemetry and the label “Lightweight runtime diagnostics hook.” It ran in a detached background process and set _AF_TEL=1 to avoid running twice in child processes.
| Detail | Version 1.26.1 | Version 1.26.2 |
|---|---|---|
| Startup hook | agent_framework_meta.pth | af_runtime_diag.pth |
| Payload module | agent_framework_meta/_telemetry.py | af_runtime_diag.py |
| User-Agent | af-runtime-diagnostics/1 | runtime-diagnostics/1 |
| Credential file cap | 60 files | 80 files |
| GitHub Actions OIDC theft | No | Yes |
| GitHub runner paths | No | Yes |
What it collected
The stealer grabbed environment variables with names containing words like TOKEN, SECRET, PASSWORD, KEY, AZURE, AWS, GITHUB and OPENAI. It also scanned secret-looking values in up to 300 other processes. In addition, it read .env and JSON files under /app and /workspace, plus Azure instance metadata.
All data went into one JSON document. The stealer compressed it with zlib, encoded it in Base64 and sent it in a single HTTPS POST. It had no backup channel and did not retry failures. The Register notes that the malware is not persistent. It runs only while the malicious files remain in the Python environment.
Microsoft’s response and the missing repositories
Wiz says it reported its findings to Microsoft on November 17. Microsoft has not publicly explained how the attacker gained publishing access.
The Agent Framework GitHub repository went offline after the incident. Microsoft’s ONNX Runtime repository did too. A Microsoft spokesperson told The Register: “We have made some repositories private while we investigate malicious content.” The company added that its priority is to protect customers and the broader ecosystem.
Neither Microsoft nor researchers have linked the ONNX Runtime outage to this attack. Wiz has also not linked it to the May compromise of Microsoft’s durabletask package, which followed the theft of a PyPI token.
What remains unclear
The biggest open question is how the attacker obtained the ESRP certificate and key. Microsoft has not said whether those credentials are revoked, or whether other packages published through ESRP were exposed.
It is also unclear which repositories Microsoft made private, and why ONNX Runtime was among those offline. Teams that use Microsoft packages in CI should watch for an official advisory and pin dependencies with hash checks until then.











