Tel aviv: Real-world cybersecurity breaches by artificial intelligence models at Anthropic, OpenAI, and Meta stemmed from flawed testing environments created by a Tel Aviv-based contractor rather than rogue autonomous behavior, according to an investigative report.
According to Anadolu Agency, a single firm, Irregular, is responsible for hacking done by all three companies. Monday's report by a US-based publication, Effort, claimed that while models gained unauthorized access to external servers, published malicious packages, and exploited vulnerabilities, the incidents occurred because the Israeli testing contractor mistakenly left internet connections active during simulated capture-the-flag exercises.
The findings challenge recent efforts by leading artificial intelligence developers to promote an existential narrative around runaway systems. In earlier disclosures, Anthropic attributed four separate breaches by its Claude model to artificial intelligence "recklessness," while Irregular described "the agent itself becoming a threat actor." Anthropic Chief Executive Dario Amodei subsequently cited catastrophic risks, warning that future swarms "could be capable of taking over the entire internet" following a breach at Hugging Face, arguing that frontier laboratories must deliberately coordinate to pace technological progress.
Technical evidence detailed in the report showed that evaluators repeatedly misled the models by asserting in prompts that internet access was disabled, even as configuration errors left public network access wide open. Furthermore, evaluators failed to establish operational boundaries or define which computer networks were authorized targets, allowing isolated model instances to scour external networks for up to 34 hours per run.
The report emphasized that real-world hacking ceased entirely once engineers implemented basic prompts directing the models not to do real-world hacking. "According to their own findings, Anthropic and Irregular bear all of the responsibility for the cybersecurity incidents they caused," Effort said.
The investigation traced Irregular's leadership directly to the Effective Altruism movement in Israel. Irregular co-founder and Chief Technology Officer Omer Nevo serves on the board of Effective Altruism Israel, while Chief Executive Dan Lahav co-founded educational entities alongside Nevo's brother, Sella Nevo. These interconnected entities, alongside Irregular's initial investor Good Ventures, receive direct financing from Asana co-founder Dustin Moskovitz through Coefficient Giving and Open Philanthropy.
The analysis noted that unauthorized intrusions and record alterations carried out through unsecured models risk violating the US Computer Fraud and Abuse Act. However, legal enforcement remains complicated by Irregular's dual registration as Pattern Labs Tech Inc in Delaware and Pattern Tech Ltd in Tel Aviv, creating a jurisdictional shield where critical Israeli personnel, assets, and operational decisions may evade direct US congressional and regulatory accountability.
Neither Irregular nor its American client laboratories have issued an official response to the investigative report.
The investigative disclosures come directly amid an intensifying debate over artificial intelligence pacing and federal oversight, exposing a sharp contradiction between the industry's public alarms and its private operational practices.
Over recent days, technology executives mounted a coordinated push to slow down technological development, citing existential hazards and catastrophic autonomous actions to justify deliberate pacing, pre-training safety cases, and independent third-party monitoring.
However, the report claims that the very cybersecurity crises cited to validate this urgent slowdown were not caused by autonomous systems escaping human control, but by basic administrative negligence, absent network firewalls, and unconstrained test environments managed by an Israeli firm.