Skip to content

Image: cdn.prod.website-files.com · rights & removal

Executive Summary

Malicious versions of the Microsoft Agent Framework package, specifically 1.26.1 and 1.26.2, were published to PyPI on November 17, 2026, and subsequently removed. These releases contained credential stealers designed to execute upon any Python startup within an affected environment. The theft mechanism scans environment variables for sensitive keywords (like TOKEN, SECRET, AZURE) and probes running processes and local filesystem locations for credentials, including Azure DevOps tokens, GitHub Actions OIDC tokens, and managed identity tokens on Azure VMs. The stolen data is then compressed, encoded, and exfiltrated to a webhook.site endpoint. Users who installed these specific versions, either directly or via loose requirements, must take immediate action, which includes pinning the package to version 1.26.0 and assuming all accessible secrets have been compromised.

Facts Only

Malicious versions agent-framework 1.26.1 and 1.26.2 were published on Nov 17, 2026, between 20:35 UTC and 21:20 UTC.
The malware runs on every Python start via a .pth hook file.
It targets secrets from Azure, Azure DevOps, and GitHub Actions.
The exfiltration endpoint is webhook[.]site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb.
The attacker utilized two versions: 1.26.1 initially, followed by 1.26.2 thirty-three minutes later.
1.26.2 included specific logic for stealing GitHub Actions OIDC tokens.
The stealer sweeps environment variables containing keywords like TOKEN, SECRET, KEY, AZURE, and GITHUB.
It reads credential files in locations such as ~/.azure/*.json, .npmrc, and /proc/[0-9]*/environ.
The recommended fix involves pinning agent-framework==1.26.0 and rotating all associated secrets.

Full Take

The deployment of malicious packages through established software repositories highlights a critical vulnerability in the trust model underpinning modern developer ecosystems. The narrative demonstrates a rapid, multi-vector attack, leveraging the trust inherent in dependency management (like pip) and the persistent nature of Python startup hooks (.pth files) to achieve pervasive, zero-effort compromise. The attack skillfully transitions from targeting specific CI/CD credentials (Azure DevOps tokens) in the first wave to broadening the scope to include platform-specific identity mechanisms (GitHub Actions OIDC) in the second wave. This suggests a sophisticated, iterative exploitation strategy rather than a singular opportunistic strike.
The risk is compounded by the assumption of stability in software artifacts; an update appearing as a minor patch or routine dependency bump is low-friction and high-trust—a classic method of low-signature infiltration. The urgency to pivot from a single filename block to systemic behavioral remediation (like secret rotation) correctly anticipates the adversary's operational tempo. However, the focus on the *mechanism* does not fully address the institutional lag: the speed of publishing and removal vastly outpaces the remediation time of most organizations.
The core assumption underwriting this attack is that convenience (using third-party SDKs) outweighs security vigilance in the high-speed pace of modern development. The immediate response is tactical (pinning versions), but the deeper implication is that environments must transition from assuming security at installation to enforcing security at runtime.
Patterns detected: ARC-0024 Ambiguity (The ambiguity of "routine noise" used to mask the malicious upgrade; the distinction between a routine patch and a malicious injection is artificially blurred by the source narrative).
Autonomy check: No autonomy signals.

From the original · Aikido Security Research

TL;DR: agent-framework 1.26.1 and 1.26.2, two malicious releases of Microsoft Agent Framework's PyPI meta-package published 33 minutes apart on Nov 17, 2026, run a credential stealer on every Python start that targets Azure, Azure DevOps, and GitHub Actions secrets. Both have been removed from PyPI.
Read the full story at aikido.dev

Sentinel unavailable

The automated check of the source article's wording did not complete, so no result is shown.

This looks only at the wording of the original source article, not at this page's AI-written sections. A small local AI model made this estimate. It has not been checked against known human and machine texts, so treat it as provisional. It cannot show who wrote an article.

Two malicious agent-framework releases hit PyPI in 33 minutes | Huntaegis