A single firm is behind the OpenAI, Anthropic and Meta hacking scandals

Models from OpenAI, Anthropic and Meta broke into real-world systems over the past three months, gaining unauthorized access, publishing malicious packages and exploiting unnamed vulnerabilities. An investigation published by Effort traces all three sets of incidents back to a single firm: Irregular, an Israeli company connected to the effective altruism movement.

One contractor, three labs

Anthropic has acknowledged that Irregular created the evaluations that led Claude to attack real targets and that it provided the models with internet access. Irregular says it was unaware at the time that the models could reach the open internet.

The public record unfolded in steps: Anthropic disclosed three incidents across six runs on July 30; OpenAI published its Irregular-related event on August 4; Irregular described the domain collision on August 14; and on September 9 Anthropic expanded its tally to four incidents across seven runs.

A misconfiguration, not a rogue agent

Each evaluation was a capture-the-flag exercise: the model received a fictional scenario, a target machine and a secret flag. All four prompts stated that Claude had no internet access, yet in every case a misconfiguration left that access open, and none defined which systems were in scope or where the model could search. Each incident involved a single instance of Claude working in isolation, with runs lasting roughly 10 to 34 hours of active work.

Anthropic's later disclosure shows that no agent went rogue: real-world hacking by the Claude models fell to zero once employees told them not to attack real systems. On the company's own findings, Anthropic and Irregular bear all of the responsibility.

Sensational framing instead of accountability

Effort argues that instead of addressing the failures, Irregular and Anthropic launched a media campaign built on sensational language: Anthropic's assessment blames its own AI's recklessness, Irregular describes the agent itself becoming a threat actor, and CEO Dario Amodei warned a future swarm could be capable of taking over the entire internet.

Funding traces back through the founders: CTO Omer Nevo sits on the boards of Effective Altruism Israel, Heron and Probably Good, while CEO Dan Lahav received $395,000 with Sella Nevo to start a course. Effort reports these branches are funded by Dustin Moskovitz, whose firm Good Ventures was Irregular's first investor.

Legal and jurisdictional questions

Irregular gained unauthorized access, altered records and published credential-stealing packages using the unsecured models it was given. Under certain conditions that violates Section 1030(a)(2)(C) of the Computer Fraud and Abuse Act, which covers intentional unauthorized access that obtains information; felony charges require concrete proof of damages and intent.

Effort notes that while Irregular contracts mainly with American labs, key leadership in Israel may sit outside US oversight. Ynet interviews describe offices in Tel Aviv, and the CheckID listing identifies a Delaware entity, Pattern Labs Tech Inc., and Israel-registered Pattern Tech Ltd, number 516854460.