Category: Security | Reading time: 3 min
Two malicious releases of "agent-framework," the PyPI meta-package for Microsoft Agent Framework, were published on November 17. Version 1.26.1 appeared at 20:35 UTC (wheel only) and version 1.26.2 at 21:08 UTC. Both have been pulled. Version 1.26.0 is clean. How the attacker obtained publishing access has not been disclosed.
What the malicious versions did
Both releases dropped .pth files (agent_framework_meta.pth and af_runtime_diag.pth) into the Python environment. Python executes .pth files at startup, so the stealer ran every time Python started, including when running pip.
The payload collected environment variables matching names such as TOKEN, SECRET, KEY, AZURE, AWS, GITHUB, HF_ and OPENAI, read /proc/*/environ, and took files including ~/.azure, .npmrc, .netrc, .git-credentials, Docker config, Hugging Face tokens and .env files.
It went further than reading stored secrets. It requested fresh Azure DevOps OIDC tokens and managed-identity tokens from the cloud metadata endpoint (169.254.169.254). Version 1.26.2 added theft of GitHub Actions OIDC tokens. Data was sent to webhook.site.
The compromise was reported by Aikido, Chainguard, Endor Labs, OX, SafeDep, Socket, StepSecurity and Wiz.
What to do
The GitHub advisory and researchers recommend removing the .pth files, pinning to 1.26.0, and rotating every credential the affected environment could reach, including cloud, CI/CD and model-provider keys.
Why it matters for agent deployers
Agent frameworks run with the credentials agents need to act: cloud identities, repository tokens and API keys. A poisoned framework release inherits all of them, and by minting new tokens it can reach beyond whatever was stored on disk.
Teams building on agent frameworks should treat them like any privileged dependency: pin versions, watch for unexpected releases, and scope the identities available to build and agent environments so one compromised package does not unlock everything. A per-agent record of which identity was used for which action makes it far easier to work out what a stolen token touched.
Source: GitHub Advisory GHSA-764m-3ppc-5m73 | The Hacker News | Wiz