PyPI Quarantines Two Microsoft AI Agent Framework Releases Over Credential Stealer
PyPI yanked two releases of Microsoft's agent-framework after Wiz and OX Security found credential-stealing code inside. The stealer fired the moment Python started, no import needed, and swept up cloud, git and CI secrets. Both versions are now unavailable, and researchers tell hit teams to rotate everything.
- Original source: thedefiant.io.
- The source published it on Nov 18, 2026 at 9:23 AM.
- No second outlet has reported it yet. Treat it as a single-source claim.
PyPI has quarantined versions 1.26.1 and 1.26.2 of Microsoft's agent-framework, a Python toolkit for building AI agents, after researchers at Wiz and OX Security flagged credential-stealing code in both. The two releases went up on Nov. 17 at 3:35 p.m. ET and roughly 4:08 p.m. ET, and were blocked from normal installation by about 4:30 p.m. ET, per Wiz's timeline. PyPI's pages for both now carry a quarantined label and point users to 1.26.0 instead.
The nasty part is when it ran. Installing either release dropped a .pth startup file into site-packages, Python's package directory, so the stealer launched the next time Python started in that environment — even if the app never imported the framework. Python's own documentation notes that executable startup lines run whether or not the module is used; the interpreter's -S option turns that processing off. OX Security said the package still pulled in the genuine framework core, so apps kept working normally while the theft happened underneath.
Wiz's payload analysis found the code grabbing secret-looking environment variables, .env files, Azure CLI credentials, git credentials, Docker configuration and Hugging Face tokens. It also reached for Azure managed-identity tokens and Azure DevOps pipeline identity tokens when the endpoints and credentials were there. The haul was compressed and sent toward an attacker-controlled webhook.site endpoint, the researchers said.
Version 1.26.2 went further, adding attempts to snatch GitHub Actions OpenID Connect tokens in workflows granting id-token: write. Those tokens can be swapped for Azure access where an Entra ID federated credential trusts the workflow — so the attack didn't hinge on finding stored passwords or API keys. Wiz also warned that 1.26.2 renamed the startup files, meaning a scan tuned to only the first release's filenames can miss the second. Wiz said it confirmed the attacker's free-tier collection endpoint captured at most 50 requests, which could still include live cloud, build-system and registry credentials.
Wiz and OX Security are telling affected teams to remove the releases and rotate any secrets that were exposed. PyPI's quarantine policy blocks installation through its index and stops maintainers from touching the releases while security review plays out. How the attackers got publishing access remained unconfirmed in Wiz's report.
TopicsMicrosoft · PyPI · Wiz · OX Security · agent-framework · Azure · GitHub Actions · Hugging Face
Source: thedefiant.io
Written by AI from verified reporting.
More stories
- Tech & AI npr.org Caught by @citra_hali6 0 viewsOpenAI's embedded safety assessors draw fire over limited access
- Tech & AI theregister.com Caught by @buta 2 viewsUK MPs push for AI watchdog that can switch off risky models
- Tech & AI x.com 1 viewJensen Huang Retains Right to Name Two Board Directors Via 3.5% Stake, Reuters Says
- Tech & AI cryptobriefing.com 1 viewSpaceX reportedly leasing $47B in Nvidia chips via Apollo-backed vehicle

