What happened
Two malicious wheels of Microsoft's official 'agent-framework' Python package (the meta-package for Microsoft Agent Framework, ~190k monthly downloads) were published to PyPI on 2026-11-17 as versions 1.26.1 and 1.26.2, roughly 33 minutes apart. Each wheel drops a .pth file (agent_framework_meta.pth / af_runtime_diag.pth) that Python executes on every interpreter start, so the credential stealer runs even if the package is never imported. It harvests environment secrets, credential files (~/.azure, .npmrc, .git-credentials, .docker/config.json, Hugging Face tokens, Azure DevOps agent creds), other processes' /proc/*/environ, Azure AD OIDC tokens, and Azure managed-identity (IMDS/App Service) tokens for ARM/Graph/Key Vault/Storage/Cognitive/AI Foundry, then exfiltrates zlib+base64 JSON to webhook.site. Version 1.26.2 expanded harvesting to GitHub Actions OIDC tokens.
Why it matters
This is a supply-chain attack directly on the AI developer ecosystem and the CI pipelines behind it. If an install pulled 1.26.1/1.26.2 on a build agent or dev host, the attacker gets Azure managed-identity/federated credentials and can pivot into cloud AI resources (AI Foundry, storage, Key Vault, Graph) — not just a single app's secrets. Because the stealer requests fresh OIDC/managed-identity tokens, rotating stored secrets alone does not close the exposure. Every AI system built or deployed from an affected environment is downstream at risk.
Attack vector
Supply-chain install/update pulls the malicious wheel; the .pth file executes the credential stealer on every Python interpreter start without the package being imported, exfiltrating Azure managed-identity/OIDC tokens and CI credentials to an attacker webhook.
Affected systems
agent-framework (Microsoft Agent Framework for Python) PyPI versions 1.26.1 and 1.26.2; clean version 1.26.0
Mitigation
Pin agent-framework to 1.26.0 (or later verified build) with hash-checking; remove 1.26.1/1.26.2; delete site-packages/agent_framework_meta.pth and af_runtime_diag.pth; isolate/rebuild affected hosts and runners; rotate Azure service-principal secrets, PATs, npm/PyPI/NuGet/Docker/Hugging Face/OpenAI tokens from a clean system; audit Azure sign-in and managed-identity usage; block outbound to webhook.site.