Subscribe
AI Agents
3 min read

Malicious Microsoft Agent Framework Releases on PyPI Harvested Cloud Credentials

Two trojaned versions of a 200,000-download-a-month AI agent toolkit ran a startup-time credential stealer aimed at Azure, AWS, GitHub and OpenAI secrets, and the attacker republished even after quarantine.

Retro brochure illustration of an open package box on a conveyor with a wire running to a cloud and keys
Retro brochure illustration of an open package box on a conveyor with a wire running to a cloud and keys

Microsoft's official Agent Framework Python package on PyPI shipped two malicious releases, versions 1.26.1 and 1.26.2, published on November 17, 2026, that carried a credential stealer. The package is Microsoft's open-source toolkit for building and coordinating AI agents, with roughly 200,000 downloads per month.

BleepingComputer reports the malware executed automatically during normal Python startup in affected environments, even when Agent Framework was never imported. The first release used a .pth file, agent_framework_meta.pth, to launch the stealer hidden in agent_framework_meta/_telemetry.py. The second restored the legitimate package initializer and moved the malicious payload into af_runtime_diag.pth and af_runtime_diag.py, a camouflage pattern that keeps the package looking intact.

A trojaned release passing a quarantine checkpoint on the distribution line

Both versions kept their dependency on the legitimate agent-framework-core 1.26.0, so the framework worked normally while the stealer ran in the background.

What it took

According to Socket, StepSecurity and Ox Security, which independently documented the attack, the stealer harvested secrets from environment variables, local configuration files and other running processes where permissions allowed. Targets included Azure, AWS, GitHub, Docker, Hugging Face and OpenAI credentials. It also attempted to obtain Azure managed identity access tokens and Azure DevOps OIDC tokens, which could extend access to further cloud resources. Version 1.26.2 added requests for GitHub Actions OIDC tokens and searches of GitHub runner directories, pointing squarely at CI infrastructure.

Collected data went to the same attacker-controlled endpoint in both releases, a webhook.site URL. That single exfiltration sink gives defenders a concrete log signature: investigators are advised to search egress traffic for the specific webhook path and separate blocked requests from completed transfers.

The attacker kept publishing rights

PyPI quarantined 1.26.1, and the attacker published 1.26.2 about 2.5 minutes later. That release was quarantined roughly eight minutes after it appeared. The quick republication means the attacker still held publishing access to Microsoft's official project after the first takedown. How the account or pipeline was compromised remains unclear, and Microsoft had not responded to BleepingComputer by publication time.

For teams that installed either version, the guidance is mechanical: isolate affected environments, stop running Python or pip until they are secured, rebuild from a known-clean base using the safe 1.26.0 release, and rotate every credential that was reachable from the compromised machine.

Why this one stings

Agent-framework tooling sits between source code, CI secrets and cloud identities, which is exactly the credential cluster a stealer wants. The incident lands the same week the European Commission confirmed a cyberattack tied to an AI penetration-testing tool, and it follows the LiteLLM pipeline compromise earlier this year. A package that orchestrates AI agents is now a supply-chain target in its own right, and the two-minute-and-a-half relapse after quarantine shows how little a single takedown buys when the publishing credential is still in hostile hands.

Sources

AI AgentsSecuritySupply Chain