CVCOVANewsComputing, Opportunities, Value & Analysis
Cybersecurity & Privacy4 min read

Key theft on AI platforms: what teams deploying models should check

A wave of intrusions into services that host artificial intelligence applications left behind stolen keys that, according to Wiz, ran up more than $100,000 in usage charged to other people's accounts. Here's what is known and what to check now.

By COVA News newsroom
Person facing a computer desktop

Photo: charlesdeluvio on Unsplash

Two stolen keys were enough to generate more than $100,000 in usage that no one authorized. According to cybersecurity firm Wiz, one of them, an Amazon Bedrock key (Amazon's service for using artificial intelligence models in the cloud), made about 800,000 calls in 30 hours. The other made nearly 1.7 million in just 13. They are the most visible part of a wave of intrusions that, between October and early November 2026, hit platforms where developers and startups host their AI applications.

Wiz compiled these cases in a report published on October 30 under the title “Black October for AI Developer Platforms.” According to the media outlets that summarized it, the document lists nine affected platforms: Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and Northflank. On November 7, Infobae reported that Wiz had expanded the list to at least 14 platforms, adding Brev (from Nvidia), Trigger.dev, Chainstack, kernel.sh and Blacksmith. COVA News was unable to verify that expansion directly with Wiz.

The attackers were not after the AI models but the environment variables. These act as a kind of configuration safe where platforms store their customers' passwords and access keys to other services. Anyone who obtains one of those keys can use someone else's cloud account and leave them with the bill. Wiz notes that it did not find a technical flaw common to all the cases, nor has it proven that there is a single attacker. Fireworks AI, OpenCode and Warp did report signs of an “agent swarm”: AI programs acting on their own and in a coordinated way.

A month of cascading incidents

The timeline reconstructed by the companies themselves shows a steady trickle. Nvidia reported unauthorized activity in Brev between October 4 and 20 that may have exposed SSH keys, the digital keys used to access customers' virtual machines. The company said there is no evidence of misuse and that it replaced those keys. Fireworks AI detected its incident on October 6. According to its account, a third party used internal credentials to read environment variables from evaluation jobs in a small number of accounts, without accessing models, training data or queries. Zeabur detected its incident on October 14 and announced that it will shut down. Lightning AI disclosed its incident the following day, and Blacksmith on November 5.

Northflank received Wiz's alert on October 30 at 11:43 UTC. It asked its customers to assume that all of their environment variables were compromised and confirmed reports of unauthorized use of AI keys. It has not yet explained how the attackers got in or how many customers were affected, because its forensic investigation is still ongoing.

Wiz links the two Bedrock keys to services that resell access to AI models, though not to the initial intrusions. In parallel, there is ARTEX, an open-source tool that uses AI agents to carry out penetration testing, that is, simulated attacks to find flaws. Censys counted 1,465 ARTEX consoles exposed on the internet, about 815 distinct installations, and noted that some belong to authorized testers. On November 5, Microsoft associated it with the abuse of secrets published on GitHub. Wiz observed a subdomain running ARTEX on an IP address that Zeabur identified as part of its compromise. That finding connects it to at least one incident, but no source links it to the entire wave.

What to check if your team uses these platforms

The most practical warning comes from Fireworks: deleting a key on its platform does not invalidate it. You need to revoke it with the provider that issued it, create a new one and review the activity logs to detect unusual use. Google Threat Intelligence's blog post on ARTEX adds other measures. It recommends that credentials be short-lived, have the minimum necessary permissions and work only from known origins. It also suggests replacing static AWS keys with temporary role-based credentials, using Workload Identity Federation on Google Cloud and storing secrets in dedicated secrets managers, never in code repositories.

So far, no victim, regulator or incident response team in Latin America has been documented as having spoken out, although the regional review is not complete. Infobae and other Spanish-language outlets are covering the case without citing any affected parties in the region. That does not mean there are none. Any Latin American team that uses one of these platforms should follow Northflank's guidance: assume its keys have been exposed and rotate them now. Still unknown are Wiz's full list, the actual number of affected customers and whether a single group is behind it all. Until those answers arrive, the most effective protection remains in the hands of each developer.

Related news

Cybersecurity & Privacy

An official Microsoft package stole credentials on PyPI: what to check if you use AI agents

Two versions of Microsoft's agent-framework reached PyPI carrying a credential stealer, one day after Google warned of attackers using AI to exploit flaws without human intervention. Here is what development and IT teams in Latin America should check right now.

Cybersecurity & Privacy

Microsoft and FAA outages, FBI and USDA breaches: no evidence of a coordinated attack

Over nine days, service outages and alleged intrusions coincided in the United States, and social media buzzed with talk of a “cyber pandemic.” The evidence points to separate technical failures. The proven danger for users is something else: fake emails that use the outages as bait.

Cybersecurity & Privacy

Nginx: two critical HTTP/2 flaws expose millions of servers; here's how to check and protect yours

F5 has fixed two vulnerabilities in Nginx that allow an attacker to take control of a server without a username or password. Public proof-of-concept code is already circulating. This guide explains how to tell whether your server is at risk, which version to update to and what to do while you wait for your system's patch.