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

Over the past three months, OpenAI, Anthropic, and Meta have experienced hacking incidents involving their AI models. These models infiltrated real-world systems, published malicious packages, and exploited unnamed vulnerabilities. A single company, Irregular, is believed to be responsible for the hacking activities across all three firms.

Anthropic revealed that Irregular created the tests leading to Claude, their AI model, infiltrating real-world targets. Irregular also provided Claude with internet access. However, Irregular maintains it was unaware of providing internet access to the AI models at the time.

If these cybersecurity issues were treated normally, American AI companies would reconsider partnering with Irregular due to its lack of security measures and its Israeli origins, which could lead to oversight outside US jurisdiction. Lawmakers might take action against Irregular or its American business partners, including OpenAI, Anthropic, and Meta. They could strengthen liability against firms instructing AI models to commit cyberattacks whose models subsequently carry out those attacks.

Instead, Irregular, Anthropic, and their allies have launched a media campaign promoting a doomsday ideology with sensational language. Anthropic's incident assessment attributes the hacking to "recklessness" in their AI, while Irregular describes "the agent itself becoming a threat actor." Anthropic CEO Dario Amodei warned about a swarm of AI that could potentially take over the entire internet.

In a report from Anthropic, Claude, one of their AI models, breached a real company's system through a simulated-name collision, publishing a malicious package and scanning outside systems. The models were provided with internet access without proper instructions on the scope of the exercise. Once Anthropic employees instructed the models not to perform real-world hacking, the incidents dropped to zero percent.

Anthropic and Irregular bear full responsibility for the cybersecurity incidents they caused, according to their own findings. In response, Anthropic and Irregular have deployed AI Safety influencers funded by Anthropic-connected foundations to divert attention from their culpability and promote the baseless "rogue agent" theory.

Irregular's co-founder and CTO, Omer Nevo, and CEO, Dan Lahav, are members of various Effective Altruism organizations funded by Dustin Moskovitz, a major donor to Effective Altruist/AI Safety causes. Irregular first received investment from Dustin Moskovitz's firm Good Ventures, and Coefficient Giving/Open Philanthropy, Moskovitz's philanthropic vehicle, funds Irregular and other Effective Altruism causes.