theaisafemode.com
AI Safe Mode

A Hijacked AI Agent Toolkit Stole Developers' Cloud Keys

A white wireless keyboard on a wooden desk; illustrative
Illustrative photograph only, not an event image. Photo by Mitchel Lensink via Wikimedia Commons; CC0 1.0. See LICENSES.md.

Two tampered releases of a big tech company's official toolkit for building AI agents hid a credential stealer that ran on startup. What developers should do now.

Attackers published two tampered versions of a big tech company’s official open-source toolkit for building AI agents. The releases looked normal and the toolkit kept working, but hidden code quietly collected cloud passwords, access tokens and other secrets from developers’ machines.[1] [2]

What happened

According to the reports, the two malicious versions appeared on Nov 17, 2026 on the public package registry of a popular programming language. The toolkit is widely used, with roughly 200,000 downloads a month.[1]

The hidden code was set up to run automatically whenever the programming language started on an affected computer, even if the developer never used the toolkit itself. It searched settings files, environment variables and running programs for keys to major cloud and AI services, code-hosting accounts and build pipelines, and sent them to a server run by the attacker.[1] [2]

The registry removed the first release, but a second tampered version appeared minutes later, which showed the attacker still had publishing access. That one was also removed shortly afterwards. It is not yet known how the attacker gained the right to publish under the official project.[1] [2]

Why it matters

AI agent projects are often connected to powerful accounts: cloud servers, paid AI services and company code. A stolen key can let attackers run up bills, read data or plant code in other projects.

Developers usually trust official packages from big publishers. This case shows that even trusted names can be used to deliver malware if an attacker gets into the publishing process.

The short version

A trusted AI agent toolkit was tampered with to steal keys. If you installed the bad versions, assume your secrets are exposed.

How to stay safe

  • Check your installed version. Security researchers advise moving back to the last clean release.
  • Rotate every key and token on any machine or build server that installed a tampered version.
  • Pin exact versions and use lock files so updates do not install automatically.
  • Give AI projects their own limited keys, with spending caps, rather than your main account keys.
  • Rebuild affected environments from a clean state instead of trying to clean them by hand.

Follow AI Safe Mode

See the post on YouTube Watch and follow AI Safe Mode on YouTube @AISafeMode and Instagram @aisafemode.

References

  1. BleepingComputer, Nov 18, 2026. Report on the compromised AI agent package and what it stole. Read the report ↗
  2. Help Net Security, Nov 19, 2026. Report on the credential stealer and the second release after quarantine. Read the report ↗
Facts come only from the published reports above and are attributed to them. AI Safe Mode doesn't name companies, products or people involved in incidents. Spotted an error? Email [email protected] and it will be reviewed and corrected. See the corrections policy.