Two compromised releases of Microsoft’s agent-framework Python package install a credential stealer that runs whenever Python starts in an affected environment, The Hacker News reported on November 18, citing security researchers. Versions 1.26.1 and 1.26.2 target stored secrets and cloud identity tokens, putting developer systems and automated build and deployment environments at risk.
The malicious releases were published 33 minutes apart on November 17, 2026. Both had been quarantined on the Python Package Index, or PyPI, by the time of the report. How the attacker obtained publishing access remains unclear.
The package, which attracts about 190,000 downloads a month, is the meta-package used to bring together components of Microsoft Agent Framework, an open-source framework for building and coordinating artificial intelligence agents in Python. The download figure describes the package’s overall use, rather than installations of the malicious releases.
The report cites findings from Aikido, Endor Labs, OX Security, SafeDep, Socket, StepSecurity and Google-owned Wiz. Organizations checking their exposure are advised to examine installations between 8:35 p.m. and roughly 9:20 p.m. UTC on November 17.
A startup hook makes every Python launch a trigger
The malicious packages place a .pth file in the root of Python’s site-packages directory. Python processes import statements in these files when its interpreter starts, allowing the attacker’s code to execute even when an application never imports Agent Framework.
Endor Labs explained in findings cited by the report that this also affects launches of pip, Python’s package installer. Simply starting Python in a compromised environment can trigger the stealer.
Version 1.26.1 installs agent_framework_meta.pth, whose single import statement loads agent_framework_meta and launches the payload. Version 1.26.2 instead installs af_runtime_diag.pth and imports a new top-level module, af_runtime_diag.py.
The payload searches environment variables using substring matches and collects credential files associated with Azure CLI (~/.azure/*.json), npm (.npmrc), pip (pip.conf), .netrc, .git-credentials, Docker (.docker/config.json), Hugging Face and Microsoft Azure DevOps agents. It also searches for .env and JSON files.
Collected information is assembled into a JSON document, compressed with zlib, Base64-encoded and sent in an HTTPS POST request to the hard-coded destination webhook[.]site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb.
Cloud identities extend the exposure beyond saved passwords
The stealer also seeks Azure DevOps OpenID Connect, or OIDC, tokens—identity tokens used in federated authentication—and Azure managed-identity access tokens. Managed identities allow cloud workloads to authenticate to other services without keeping a password in application code.
SafeDep’s analysis, quoted in the report, explains that the payload can request fresh credentials from the platform. In an Azure DevOps pipeline, it uses SYSTEM_ACCESSTOKEN and SYSTEM_OIDCREQUESTURI to request a federated token for api://AzureADTokenExchange.
On Azure compute, the payload requests managed-identity access tokens through IDENTITY_ENDPOINT or MSI_ENDPOINT on App Service and Functions, and through 169.254.169[.]254 elsewhere. It also retrieves Azure instance metadata, retaining up to 8,000 characters of the response. The metadata service supplies information about a virtual machine, including its region and tags, alongside an endpoint for requesting identity tokens.
Version 1.26.2 adds collection of GitHub Actions OIDC tokens and expands the patterns used to find credential-related environment variables. That makes workflows deploying from GitHub Actions to Azure a specific response priority.
The ability to request tokens means rotating stored secrets alone does not cover the exposure, SafeDep warned. The practical reach of a stolen managed-identity token depends on the resources that identity can access. In a separate October 2023 analysis, Microsoft described an unsuccessful attempt to abuse a SQL Server host’s cloud identity and recommended granting managed identities only the permissions they need.
OX Security suggested that attackers appeared to have hijacked the wheel creation process—the process that builds Python distribution packages—to insert the malicious logic. Earlier in May 2026, attackers associated with TeamPCP compromised three releases of Microsoft’s durabletask Python client to distribute the Mini Shai-Hulud worm, the report noted.
Response for affected environments
The response steps published with the report call for administrators to:
- Isolate affected hosts and automated job runners.
- Pin
agent-frameworkto 1.26.0, identified as the most recent safe version, and restore from a clean environment. - Search proxy and DNS logs for activity involving
webhook[.]site, checking available request records for the full destination above. - Rotate exposed credentials from a clean system.
- Audit Azure identities and review sign-in and activity logs.
- Audit GitHub Actions workflows that deploy to Azure.




