Skip to main content
QUIETLYTIC
Supply chain

Two agent-framework releases on PyPI shipped a CI credential stealer

PyPI agent-framework 1.26.1 and 1.26.2 ran a credential stealer on install, aimed at cloud, CI and GitHub Actions tokens. Rotate secrets if you pulled them.

Two agent-framework releases on PyPI shipped a CI credential stealer — Supply chain
Category
Supply chain
Severity
Critical

Two releases of agent-framework, a Python package for building AI agent applications, carried an information stealer when they were published on 17 November 2026. OSV and the OpenSSF Malicious Packages project report that versions 1.26.1 and 1.26.2 run hidden code the moment the package is installed and send secrets from the host to an outside server. Both versions have been pulled from PyPI, but any machine or build job that resolved them before they were removed should be treated as compromised.

The package is a legitimate project, not a typosquat. That matters: teams that track the latest release, or that rebuilt containers and CI images on the evening of 17 November UTC, could have taken the bad version with no unusual name to warn them.

What the malicious releases did

According to the OSV record, version 1.26.1 appeared at about 20:35 UTC as a wheel only, and 1.26.2 followed at about 21:08 UTC. Release 1.26.0 is the last clean one, and everything earlier is unaffected.

The added code does not wait for an application to import the library. Each wheel drops a Python startup hook, a kind of file the interpreter processes every time it launches, so the stealer starts on any Python run in that environment after installation. It then forks into a quiet background process and suppresses its own errors, which leaves little visible trace in build logs. Apart from these additions, the record says the wheels match 1.26.0 and keep the same pinned core dependency, so the package keeps working normally and gives users no functional reason to suspect it.

The second release was not a fix. It renamed the added files, dropped the hook from the package’s own initialisation code so the startup file became the only trigger, and broadened what it collected.

What it was built to take

The record describes a collector aimed squarely at developer workstations, cloud workloads and build runners. It gathers:

Target Detail from the record
Host context Hostname, user name, working directory and command-line arguments
Environment secrets Variables whose names suggest tokens, passwords or keys, plus those tied to Azure, AWS, GitHub, Hugging Face, OpenAI, npm, pip and CI systems; also the environment of other running processes
Credential files Azure CLI configuration, npm and netrc files, stored Git credentials, Docker registry config, Hugging Face tokens, Azure DevOps agent credentials and .env files
Pipeline identity Azure DevOps access and OIDC request tokens
Cloud workload identity Azure managed-identity tokens requested from the instance metadata service
GitHub Actions (1.26.2) Actions OIDC request tokens and the Actions runtime token, with GitHub-hosted runner home paths added and the file cap raised from 60 to 80

The data is compressed, encoded and posted over HTTPS to an external collection endpoint that is the same in both versions. Because the harvest covers the environment of other processes, a single infected interpreter on a shared build host can expose secrets that belong to unrelated jobs.

Are you affected?

Exposure requires that version 1.26.1 or 1.26.2 was actually installed. Check in this order:

  1. Lockfiles and pins. Search requirements*.txt, poetry.lock, uv.lock, Pipfile.lock and similar files for agent-framework at either bad version. A pin at 1.26.0 or lower protects you; a loose range such as >=1.26 did not during the window.
  2. Installed environments. Run pip show agent-framework or pip list in each virtual environment, container image and runner image built since the evening of 17 November UTC.
  3. CI history. Review dependency-install logs for pipeline runs in that period. Ephemeral runners leave no package behind, so the logs may be the only evidence.
  4. Package proxies. If you mirror PyPI through an internal registry, check whether either version was cached, and purge it.

What to do now

  1. Rotate first, clean second. Assume every credential reachable from an affected host or job has been read. Rotate cloud keys, Azure service principals and managed-identity permissions, GitHub and Azure DevOps tokens, registry passwords, npm and PyPI publishing tokens, and model-provider API keys. The record specifically calls out CI and GitHub Actions secrets.
  2. Review federated trust. Stolen OIDC tokens are short-lived, but check cloud audit logs for role assumptions or token exchanges from your pipelines that do not match a real build.
  3. Remove and pin. Uninstall the package, reinstall 1.26.0 or a later release confirmed clean by the project, and rebuild affected images from scratch rather than patching them in place. Look in each environment’s site-packages directory for startup files that do not belong to a package you recognise.
  4. Harden the next install. Require hash-checked installs from lockfiles, and consider a short delay before newly published versions can enter production builds.

What remains unclear

The record does not say how the malicious versions reached PyPI, whether through a stolen publishing credential, a compromised release pipeline or some other route. It does not tie the activity to any known campaign, and it gives no count of how many installs occurred before removal. Whether the project will publish its own incident account, or whether other packages were tampered with in the same way, is also unknown. We will update this article if the record changes.

Frequently Asked Questions

Which versions of agent-framework on PyPI are malicious? Versions 1.26.1 and 1.26.2, both published on 17 November 2026 and since removed. Version 1.26.0 and earlier are not affected.

Does the malware run if my code never imports agent-framework? Yes. The releases add a Python startup hook, so the stealer runs on any interpreter launch in that environment once the package is installed.

What should I do if a build installed one of the bad versions? Treat the environment as exposed: rotate every credential it could reach, including CI and GitHub Actions secrets, then remove the package and rebuild from a clean image pinned to a safe version.


Contains data from the OpenSSF Malicious Packages dataset via OSV.dev, licensed under Apache-2.0.

Report an error

Found a factual error, an outdated figure, or a broken source link? Let us know and our editorial desk will review it.


Sources

  1. OSV — MAL-2026-20101: Malicious code in agent-framework (PyPI)

Related intelligence


Analyst tools