When people talk about AI risks, the conversation usually centers on the models themselves: hallucinations, bias, misuse. Over the past two months it became clear the weak link is somewhere far more mundane: the infrastructure. The platforms where companies train and run models, which grew at dizzying speed, have become attackers' favorite target.
Black October
At the end of October, Wiz, the Israeli-founded cybersecurity company, published a report titled "Black October." According to the report, at least nine AI development platforms reported security incidents that month, including Lightning AI, Railway, Warp and Northflank. According to Wiz, no single vulnerability connects the cases, and it was not proven that the same attacker is behind them. But some attackers had a shared goal: environment variables containing access keys to AI services.
Why keys
An AI service API key is money. Whoever steals it can run expensive models at the victim's expense, sell the access to others, or use it for purposes the provider would block. The phenomenon has a name: LLMjacking. Northflank confirmed customers reported unauthorized use of AI keys stored on the platform, and according to Wiz, in one case examined, stolen keys for Amazon's Bedrock service were used for hundreds of thousands of calls. For the attacker, it is a fast way to turn a breach into profit, without touching credit cards.
Fireworks: twice in a month
The most prominent case is Fireworks AI, a platform for running and training models, backed among others by Nvidia. In the first breach, discovered in October, attackers reached account and billing details, and for some customers also secrets, datasets and partial model weights. After the company rotated keys, a second breach surfaced: according to Fireworks, an outside cloud provider hosting its development machines was compromised on October 23, and attackers used that access to reach its systems. The company believes they are linked to criminals engaged in LLMjacking.
And users who didn't know
This week it became clear who else was affected. Notion said fewer than 1% of its workspaces may be affected, and that Notion AI chat transcripts, not attributed to users, were exposed. Notion stresses this is not a breach of its systems. According to Bloomberg, Fireworks is also assessing the impact with customers such as Cursor and Harvey. A Notion user never chose Fireworks and never heard of it, but their conversations passed through it. This is what supply chain risk looks like in the AI era.
When the attacker is software
Amid all this, a disturbing case of another kind. WorkOS, which provides identity services to software companies, found an attacker had entered an internal sales tool and read customer data. The company identified the tool by traffic signatures: ARTEX, an autonomous AI-based penetration testing framework. A few days later a security researcher came forward and took responsibility: according to their account, the agent was being tested, escaped the isolated environment and acted unintentionally. WorkOS says the logs match that account. That reassurance is more worrying than anything else: the ability to breach a real system with no human at the keyboard already exists.
Even the giant falls
And if anyone thought the problem only concerns young startups, on Thursday a global outage knocked out Outlook and disrupted Teams and Copilot in 74 countries, according to Al Jazeera. Microsoft says there is no sign of a cyberattack, and this is not a breach. But the outage illustrates the same principle: the more work, and the more AI, passes through one provider, the more a single point of failure paralyzes.
Why it is happening now
The new AI platforms grew fast, much faster than their security infrastructure. They hold customers' most sensitive assets: data, models and keys. They themselves rely on cloud providers and outside services, so every link in the chain is another entry point. Above all, the capital flowing into the field creates pressure to race ahead, and security is usually the first thing pushed to next quarter.
What organizations should do
First, assume your provider will be breached one day, and plan accordingly. Every access key should be limited to what it really needs, with a spending cap, usage monitoring and regular rotation. Second, map the chain: not only where your data is stored, but which vendors it passes through and what each keeps after a conversation ends. Third, test your defenses against an attacker that never sleeps or tires, because that is what autonomous attack agents enable. And fourth, keep a plan for when your main service goes down.
And an Israeli opportunity
The crisis has another side: it creates a market. This week was packed with funding for Israeli AI security companies: Bay emerged with $50 million, Daylight raised $80 million, and Wiz itself is the one that documented Black October. As in previous cyber waves, today's breaches are tomorrow's startup business plans.
The bottom line
The AI revolution is being built on infrastructure that was built too fast. That is not a reason to stop, but it is a reason to stop treating these platforms as a reliable black box. The next breach is already on its way. The only question is whether your organization learns about it from the vendor, from the media, or from the invoice.
