First depth founders (from left to right): Qasim Mithani, Daniele Perito and Andrea Michi.
Deep down first
Andrea Michi spent nearly seven years at Google’s Deepmind developing artificial intelligence models that improved energy efficiency across wind farms and data centers. He realized that the most effective way to build AI that focuses on solving big problems is not to start with LLMs, and then prepare them to go to work: it is to build models from the ground up with a specific purpose in mind.
In 2024, Michi left Deepmind and founded Depthfirst to apply that thinking to cybersecurity, creating AI models that automatically find vulnerabilities and suggest fixes. The company is developing what it calls “General Security Intelligence,” a game about general artificial intelligence, meaning AI capable of general tasks like humans. His hope is that Depthfirst will find and fix vulnerabilities in any kind of software before a malicious, intelligent AI can exploit them.
“We want to make sure that these complex and sophisticated systems can support a world where AI’s hacking capabilities are at their peak and it means staying one step ahead of attackers,” said Michi, who is CTO of Depthfirst. Forbes.
“The agent tries to find the most exploitable vulnerability.”
The stakes are high, but investors are buying. On Tuesday, the two-year-old company announced an $80 million raise at a $580 million valuation, just two months after a $40 million Series A round. CEO and co-founder Qasim Mithani, former head of infrastructure security at the $134 billion data analytics company Databricks, says. Forbes the company is growing rapidly. Its revenue has increased by 300% over the previous two quarters and it has more than 20 customers, from multinationals to AI startups like the $6.6 billion coding platform Lovable. Next to Michi and Mithani, the third co-founder is Daniele Perito, who was the director of network security at Jack Dorsey’s payments company.
Arsham Memarzadeh, Meritech’s lead Series B investor, believes that Depthfirst is well positioned to dominate the $18 billion application security market, where companies are focused on protecting applications rather than servers, computers or industrial systems. Just as Wiz quickly overtook cybersecurity industry giants like Symantec and McAfee to dominate the cloud security space before Google acquired it for $32 billion, Memarzadeh believes Depthfirst has the leadership talent and technology to make that breakthrough for its industry. “In cybersecurity, I think it’s very difficult to be a big sustainable business unless you have one of the core threats,” he says. If you don’t, you end up being taken to a site and not a product.
The startup is launching its first in-house version this week, dfs-mini1, aimed at detecting security vulnerabilities in crypto smart contracts. Like other Depthfirst models, the company has trained this AI through reinforcement learning, where agents learn through trial and error, rather than dealing with large volumes of recorded data. It did so by placing agents in places where they were tasked with finding vulnerabilities in smart contracts. Every time the AI got it, it got a reward. “We started with smart contracts because that is one area where vulnerability can cause disaster, which can lead to loss of income,” says Michi.
Other designs will follow to cover some of the above, says Mithani. “Our agents are very serious,” he says.
He says building models from scratch rather than adapting existing ones is cheaper because Depthfirst can control the energy needed to train them. The company also doesn’t charge the likes of OpenAI and Anthropic to use their models, though it does have to pay for infrastructure to run its technology.
With its funding increase this year, Depthfirst plans to build AI that can handle many of the tasks that cyber security teams do today. That doesn’t mean Depthfirst plans to cut security staff. Michi says: For me the question is: can we add these people now to help face a bigger threat than we have now?”
MORE FORBES
#Meet #Billion #Startup #Build #Models #Fight #Artificial #Attackers