The US cyber agency CISA has added a flaw in GPUStack, open-source software that organisations use to manage the GPU servers that run large AI models, to its list of vulnerabilities known to be used in real attacks. It did so on 9 November. The flaw, CVE-2026-102895, is rated critical (9.8 out of 10). GPUStack’s own advisory says it lets an attacker gain full administrator access over the network without signing in, and that it has been “observed exploited in the wild”. A fix has been available since version 2.2.0, released in June. A second flaw in the software’s single sign-on login, CVE-2026-102894, is fixed in version 2.2.2; no attacks on it have been reported.
Why it matters: the servers behind AI models are valuable targets. They hold costly computing power and often the keys to cloud accounts, and an attacker with admin access can change which models run and use the machines for their own ends. Securing AI is not only about tricking chatbots; the plumbing underneath needs patching too.
What to do: If your organisation runs GPUStack, upgrade to 2.2.2 or later now and, as the advisory recommends, keep its API off untrusted networks. Everyone else needs to do nothing.