Yesterday at 20:35 UTC, two malicious versions of Microsoft’s official agent-framework Python package showed up on PyPI. They were live for 55 minutes before removal. If you installed agent-framework==1.26.1 or 1.26.2 during that window — in a local environment, a CI/CD pipeline, or a Docker build — treat every credential on that system as stolen and start rotating now.
The 55-Minute Window
The last clean release was 1.26.0, published November 12. On November 17, the attack played out fast:
- 20:35 UTC: Version 1.26.1 uploaded to PyPI
- 21:08 UTC: Version 1.26.2 uploaded (33 minutes later, as 1.26.1 was being quarantined)
- ~21:30 UTC: Both versions removed from PyPI
The attacker adapted in real time. When PyPI began quarantining 1.26.1, a second, re-obfuscated version was ready two minutes later. That’s not a script kiddie — that’s a prepared operation.
Why .pth Hooks Are Worse Than a Bad Import
Both versions used Python’s .pth startup hook mechanism to run the credential stealer. This is what makes the attack nastier than a typical malicious package.
Normally, a malicious package needs you to import it before anything executes. A .pth file bypasses that entirely. Python executes any import line in a .pth file on every interpreter start — including when you run pip, pytest, or any other Python tool. You don’t have to touch agent-framework in your code. Installing the package is enough.
Version 1.26.1 routed the stealer through agent_framework_meta.pth → agent_framework_meta/_telemetry.py. The _telemetry.py name is deliberate social engineering — it looks like normal observability code. Version 1.26.2 switched to af_runtime_diag.pth and expanded scope to include GitHub Actions OIDC token collection. JFrog Security Research published a full technical breakdown of both versions.
Traditional static analysis tools that scan for suspicious import statements can miss this entirely.
What Was Taken
The credential stealer targeted a broad surface area:
- Azure CLI token caches and Azure DevOps tokens
- GitHub Actions secrets,
GITHUB_TOKEN, and OIDC tokens - Config files:
.npmrc,.git-credentials,.docker/config.json,pip.conf - Hugging Face tokens and OpenAI API keys
- General environment variables matching secrets or token patterns
Everything was exfiltrated to a webhook.site endpoint — a free HTTP receiver that leaves minimal forensic trace. Wiz published an impact analysis focused on Azure credential exposure.
How to Check if You’re Exposed
Do not run pip show agent-framework or pip uninstall first on a potentially affected machine — both start Python, which triggers the stealer if it’s present.
Start with your lockfiles (safe, no Python invoked):
grep -rnE 'agent[-_]framework.{0,20}1\.26\.[12]\b' \
--include='*.txt' --include='*.lock' \
--include='*.toml' --include='*.cfg' .
Then check for the malicious .pth artifacts directly in your site-packages:
find /path/to/site-packages \
-name "agent_framework_meta.pth" -o \
-name "af_runtime_diag.pth" 2>/dev/null
Also check CI logs from November 17 for lines like Installing collected packages: agent-framework-1.26.1.
Remediation: If You’re Affected
If you find evidence of 1.26.1 or 1.26.2:
- Manually remove the
.pthhooks before running pip:agent_framework_meta.pth,af_runtime_diag.pth, and their associated directories - Pin to
agent-framework==1.26.0with--require-hashes - From a separate, clean machine: rotate every credential the affected environment could have read — GitHub tokens, Azure DevOps service connections, Hugging Face tokens, CI/CD secrets
- Audit Azure Entra ID sign-in logs for unexpected activity on November 17–18 from unfamiliar IPs
- Review and remove unexpected federated identity credentials on Azure apps and managed identities
StepSecurity published CI/CD-specific guidance if your pipelines are the concern.
Second Time This Year for Microsoft’s Python Stack
This is not the first time a Microsoft AI Python package has been hit. In May 2026, TeamPCP compromised durabletask — another Microsoft Python SDK — by extracting a PyPI upload token from GitHub Actions secrets and publishing three malicious versions. Both incidents follow the same pattern: a valid PyPI upload credential, a package used by AI developers with elevated cloud access, and a payload designed to harvest exactly the credentials those developers hold.
The pattern is wider than Microsoft: 56 of 106 tracked supply chain campaigns in 2026 H1 targeted AI, LLM, or agent tooling. Developers building on AI frameworks tend to hold powerful cloud credentials, which makes their dependency trees a high-value attack surface.
Until Microsoft publishes a post-mortem explaining how the upload credential was obtained, pin your versions, enable hash verification, and check your lockfiles after every CI run. The supply chain is the attack surface now.

