Irregular raises $80 million to secure frontier AI models

On Wednesday, IRRGULAR Security Company announced $ 80 million in new financing on a tour led by Sequoia Capital and Redpoint Ventures, with the participation of WizAf Rappaport CEO. A source close to the deal said that the irregular tour of $ 450 million.

“Our point of view is that it is soon, a lot of economic activity will come from the human interaction on the AI ​​and the AI-AI’s interaction,” co-founder Dan Lahaf told Techcrunch.

Previously known as Labs Pattern Labs, it is already an important player in artificial intelligence assessments. The company’s work was cited in security assessments Claude 3.7 Sonata Besides Openai’s O3 and O4-MINI models. In general, the company’s framework to register the ability to detect weakness on the model (Solutions solutionsIt is widely used in industry.

Although irregular has done a great job in the current risks of models, the company collects donations with the eye towards something more ambitious: discovering the risks and emerging behaviors before they appear in the wilderness. The company built a detailed system of simulation environments, allowing an intense test of a model before its release.

“We have a complex network simulation as we take over the role of the striker and defender,” says Amer Nevo, co -founder. “So when a new model comes out, we can see where the defenses are held and where it does not do it.”

Security has become an intensive focus point for the artificial intelligence industry, as the potential risks posed by border models with more risks. Openai Reforming internal security measures This summer, with the eye towards the spying of potential companies.

At the same time, artificial intelligence models Increasingly In finding weaknesses in software – a force with serious effects on both attackers and defenders.

For irregular founders, this is the first of many security headaches caused by the increasing capabilities of large language models.

“If the goal of the border laboratory is to create more sophisticated and increasingly capable models, our goal is to secure these models,” says Lahaf. “But it is a moving goal, so by nature there is a lot, and more work that must be done in the future.”