Why Lightspeed and Wiz’s Assaf Rappaport bet $37 million on an AI-powered cyberattacker

After frontier AI models exposed thousands of previously unknown vulnerabilities, startup A is building autonomous defenses designed for an era of AI-native attackers.

Mock cyber attacks used to cost about $30,000, take two weeks, and tell companies what their defenses looked like six months ago. Then Anthropic’s Mythos showed up, exposing thousands of zero-day vulnerabilities.

A (yes, as in the letter) uses AI to continuously break into its own customers’ systems—finding real attack paths and fixing them—before actual hackers get the chance. The New York-based autonomous offensive security startup, emerged from stealth with $37 million in funding from Lightspeed Venture Partners, Cyberstarts, and angels including Wiz CEO Assaf Rapaport and Cyera CEO Yotam Segev, Fortune learned exclusively.

The timing is hard to miss. Anthropic’s Mythos model—unveiled April 8 in a restricted release—autonomously surfaced previously unknown security flaws across every major operating system and browser. This handed unsophisticated attackers a preview of what’s coming. “Today every teenager, every grandmother can take a laptop, access those frontier models, and do things that were once reserved for nation-state actors,” A CEO and cofounder Yossi Torati told me. “We will face a meltdown in the industry.”

Torati spent six years at Israeli incident response firm Sygnia helping organizations investigate and recover from major hacks and responding to breaches across roughly 40 countries. His cofounders, Omer Gull and Yuval Itzchakov, come from similar cyber backgrounds from Check Point, Hunters, and IDF Unit 8200. 

A’s founding moment was when Torati saw attackers begin to use AI and realized that most of what the industry had built was about to become obsolete. “This is a fundamental revolution of how attackers operate and how defenders need to respond,” he said.

Instead of flagging risk hotspots or handing companies a severity score—A runs the full cyber offensive lifecycle autonomously, acting like an AI-powered attacker to find weaknesses, then helping fix them before a real attacker can. In one early proof-of-concept, A turned up 1.2 million sensitive customer records, including Social Security numbers, that had been sitting exposed for seven years, undetected by the customer’s own team and existing tooling. Torati didn’t disclose the customer’s name.

That’s the white space Lightspeed partner Guru Chahal says his firm was hunting for. “The gap between the time where an issue is discovered to the time you get hit with it is minutes,” Chahal told me. “CISOs are terrified of this right now.” He says the old model—paying for a manual simulated cyberattack once every six to 12 months—is effectively dead. “In those six months, someone’s going to walk away with all your crown jewels. And so what’s the point, even?”

The market A is chasing—continuous threat exposure management—was estimated at $2.7 billion in 2025 and is projected to reach $7 billion by 2033. The competitive field includes well-funded AI-native players like XBOW, which raised $120 million at a $1 billion-plus valuation in March.

Chahal, for his part, isn’t hedging. “I genuinely believe some of the largest companies are going to get built in this category,” he told me. 

The capital will go toward scaling go-to-market and building out the product roadmap. Current A customers are all enterprises—finance, healthcare, critical infrastructure, and technology—though Torati declined to give a number. 

As for the name, Torati says it represents going back to first principles—the attacker’s perspective, the root of what cybersecurity was always supposed to be. Being first in the alphabet, he acknowledged with a laugh, doesn’t hurt either.

See you tomorrow,

Lily Mae Lazarus
X: @LilyMaeLazarus
Email: lily.lazarus@fortune.com
Submit a deal for the Term Sheet newsletter here.

Joey Abrams curated the deals section of today’s newsletter. Subscribe here.

This story was originally featured on Fortune.com