The US Cybersecurity and Infrastructure Security Agency added four vulnerabilities to its Known Exploited Vulnerabilities catalog on 9 November, citing evidence of active exploitation. The most consequential for AI teams is GPUStack: CVE-2026-102895, CVSS 9.8, an improper-authentication flaw that lets an unauthenticated attacker take administrative control of a GPUStack deployment — changing users and models, minting persistent admin API keys, reading worker and cloud-provider secrets, and executing code through POST /v2/inference-backends/from-yaml.

GPUStack is an open-source cluster manager for GPU fleets and model serving — the layer between a company's hardware and its model endpoints. The credentials held there are exactly the ones that unlock cloud accounts and inference infrastructure. The flaw was fixed in version 2.2.0 in July 2026. Anyone still on an older release is running an internet-reachable admin bypass whose exploitation is now documented in the wild.

The other three

- CVE-2026-116779, SignalWire FreeSWITCH, 9.8: code injection. Any authenticated account can execute arbitrary code, read credentials and control calls. Fixed in 1.11.3 and later. - CVE-2026-123218, IBM Aspera Faspex 5.0.0 to 5.0.19, 9.8: incorrect authorization. A crafted message bypasses authorization and grants admin rights. Fixed in 5.0.19 PL1 or later. - CVE-2026-123158, OpenSSH ("SSHole"), 8.1: on systems using musl libc without PAM, with at least one account holding a SHA-256 or SHA-512 password hash, accounts with no password set accept any password between 257 and 1024 bytes long. Fixed in OpenSSH 10.6p2.

The OpenSSH bug is the same one reported earlier this week. GreyNoise observed the first exploitation attempt on 7 November at 00:46 UTC; it has confirmed 18 attack chains from six source IPs, including queries to cloud metadata services, attempts to read password hashes and root SSH keys, and edits to authorized_keys. In at least one case the Python-based command-and-control agent "TAC Agent" was deployed. watchTowr also saw botnet and cryptominer activity.

Three days

The Hacker News reports that federal civilian agencies have until 12 November to remediate — three days after the catalog entry. CISA's alert points to Binding Operational Directive 26-04, which also requires agencies to check whether systems were compromised before the patch was applied. As usual, the agency published no details on scale or actors, and no indicators of compromise.

What administrators should do now

Inventory GPUStack, FreeSWITCH, Aspera Faspex and OpenSSH, including versions that are not in the CMDB. For internet-exposed, unpatched instances, assume compromise: review and rotate admin accounts and API keys, reissue worker and cloud-provider secrets, audit authorized_keys and new SSH keys, check inference-backend YAMLs for foreign entries, and take management interfaces off the open internet.

Why this is more than patch Tuesday

The KEV list has become a map of AI infrastructure risk. That a flaw fixed in July is considered actively exploited in November says less about the maintainers than about the gap between patch availability and patch installation. And four critical bugs in one batch shows how much of the stack sits one unauthenticated POST away from total control.