Report Alleges Israeli Firm Behind AI Hacking Incidents Involving Anthropic, OpenAI and Meta Models

A report covered by The Cradle alleges that an Israeli technology firm is responsible for a series of artificial intelligence hacking incidents involving models from Anthropic, OpenAI and Meta. The account does not describe scattered misuse by anonymous users; it points at one specific company as the actor behind the incidents.

The most consequential detail in the report is the list of model providers it names: Anthropic, OpenAI and Meta. Those three companies build and distribute some of the most widely used general-purpose models on the market, so any organized effort to attack or abuse their systems carries consequences well beyond a single product and pulls several security teams into the same problem at once.

What the report describes is not a flaw in one model but a pattern in which an outside organization targets multiple vendors. Once model attacks can be bought or outsourced like a service, the center of gravity in defense shifts away from patching individual vulnerabilities toward identifying the funding, organization and customer chain behind the attacker.

For the model providers, a story like this moves pressure toward abuse detection and account governance: spotting anomalous usage patterns, removing accounts that try to bypass safety policies, and restricting access for high-risk use cases. Work that used to sit quietly inside security teams is now increasingly tied to corporate reputation and to conversations with regulators.

The three companies also differ in how they distribute and gate their models, which means the same kind of abuse is not equally traceable across all of them. Even when an attacker uses a single method, the evidence each vendor can gather and the action each can take will not line up, making cross-company forensics a hard problem.

At the industry level, allegations of this kind tend to trouble vendors more than a single software bug, because they raise a question of responsibility: when a third party turns a model into a weapon, the platform has to show it exercised reasonable oversight rather than issuing clarifications after the fact.

There are three things to watch next. First, whether the report produces verifiable technical detail and an evidence chain. Second, whether Anthropic, OpenAI and Meta respond or disclose enforcement action. Third, whether the episode pushes providers to tighten external access and abuse-reporting channels.