NVIDIA Brev Breach Exposed SSH Keys Over 16 Days

NVIDIA has confirmed unauthorized activity on its Brev cloud development platform spanning 16 days, from October 4 to October 20, and has replaced SSH keys that may have been exposed to attackers. The company says it found no evidence that the keys were actually misused.

The breach, first reported through a notification sent to affected customers, targeted the platform developers use to spin up and manage virtual machines in the cloud. Brev relies on SSH keys to give customers access to those instances. A stolen key can hand an attacker control of a virtual machine, depending on permissions and other security controls — the developer equivalent of handing over the remote desk password.

"We are writing to inform you that NVIDIA Corporation has experienced a security incident that may have affected the security of your data," the notification reads. "Our analysis determined that your information from our Brev network may have been affected. This SSH key has since been rotated, and your instances have been updated with the replacement key."

What NVIDIA did not say

The company stopped short of explaining how the attackers got in. It has not disclosed whether customer data or workloads were accessed, nor whether anyone attempted to use the compromised keys before they were rotated. The gap is notable. Sixteen days is a long exposure window for credentials of this kind, and "no evidence of misuse" is not the same as evidence of no misuse.

For context: SSH is the protocol administrators and developers use to connect securely to remote machines, transfer files, and run commands without physical access. It is standard plumbing for cloud development platforms, which is precisely why attackers hunt for it.

How bad was October for AI platforms?

The NVIDIA incident fits a broader pattern. On October 30, cybersecurity firm Wiz published a blog post it called "Black October" for AI developer platforms, identifying at least nine platforms hit by security incidents during the month. The list includes Zeabur, Fireworks AI, Lightning AI, Opencode, Warp, and Northflank.

Wiz researchers described the incidents as recurrent and targeted rather than opportunistic. "All public reports highlight targeted exploitation, explicitly hunting stored environment variables, with a focus on those granted AI access," they wrote.

The numbers behind that hunting are concrete. Wiz documented two cases involving compromised Amazon Bedrock credentials. In one, attackers generated roughly 800,000 API calls over 30 hours. In the other, a single key was used for about 1.7 million calls over 13 hours. Combined, the usage would have cost $100,000 at list prices.

Who else is affected?

The credential problem extends well beyond NVIDIA. Earlier this month, major US banks including JP Morgan, Morgan Stanley, and Bank of America restricted AI agents from accessing customer accounts. Their concern: automated assistants could drift from their instructions and initiate unauthorized approvals.

The attack side is equally active. The CyberCartel hacking group has used Anthropic's Claude and xAI's Grok to conduct attacks on more than 400 Latin American institutions, mostly banks and mobile payment services, after obtaining customer login credentials. Separately, China-linked attackers have been accused of deploying AI agents in a ConnectWise supply chain attack that exposed roughly 2,250 Microsoft 365 tenants. In some of those cases, attackers reportedly reached privileged access in under three hours.

What should affected customers do?

Rotating a key is a start, not a finish. Bradford Regeski, a threat intelligence analyst at RH-ISAC, argues that updating compromised software may not be enough to evict an attacker who has already established a foothold. He recommends that customers in SSH-related incidents "isolate hosts that installed the compromised extension during the exposure window," remove unauthorized SSH keys and persistence mechanisms, and rotate credentials and sessions.

For Brev users, that exposure window ran from October 4 to October 20. NVIDIA's rotation is done; the verification work now sits with the customers. Expect scrutiny of developer-platform credential handling to intensify as Wiz's October tally circulates, and as attackers continue to price stolen AI access at six figures per weekend.