The Pentagon Is Using This $1.2 Billion Startup’s AI To Automate Cyberwarfare

Earlier this year, Joe Lin, cofounder and CEO of AI startup Twenty, met with venture capitalists Vinod Khosla and Jon Chu in San Francisco. After an hour and a half of proselytizing about the company’s plans to automate hacking campaigns for the Pentagon, he’d convinced Khosla Ventures to put $30 million into the business. It’s now valued at $1.2 billion with $168 in total funding, less than a year after Forbes first reported it was coming out of stealth with backing from the CIA’s investment arm In-Q-Tel.

Twenty’s software uses AI to run hacking campaigns simultaneously, automating the gruntwork of cyberwarfare. That includes identifying targets and carrying out reconnaissance, or deciding what to do once that target has been compromised. Before next-generation AI was available, a cyber campaign against a single target could take months or years, Lin says. Now it’s much faster.

“We’re trying to industrialize cyber warfare,” Lin says. “We're going from current processes that are highly artisanal to ones that can scale out industrially, where you're able to simultaneously prosecute hundreds of targets versus one at a time.”

Khosla, who ranked No. 1 on this year’s Midas list of top venture capitalists, supports using artificial intelligence in warfare. When Anthropic, the $965 billion maker of Claude, had a public fracas with the Pentagon over use of its models, Khosla said he disagreed with the company’s stance, urging the U.S. to invest in autonomous AI weapons. Now, Khosla is putting his money where his mouth is with an investment in AI-powered cyber offense.

“AI changes what is now possible on the offensive side, collapsing the time and effort it takes to run complex cyber operations,” he said in an emailed statement. “What once required weeks of specialized human work can run at software speed… Twenty has the technical depth and the customer trust to do that.”

Twenty’s only customer is the Pentagon, but public records show only a handful of contracts, which isn’t surprising because highly sensitive or classified work is not made public. The most recent public contract is from December, signed with the Air Force Office of Special Investigations, worth as much as $640,000. Twenty is providing the agency with cyber tools to target so-called “advanced persistent threats,” or APTs, an industry term for state-sponsored hackers. As Forbes reported last year, it also signed a deal with the U.S. Cyber Command, worth up to $12.6 million. The startup declined to provide complete revenue figures.

The Department of War isn’t currently buying from Anthropic after the White House declared the company a supply chain risk (Anthropic sued, and the legal battle is ongoing). But its cyberwar efforts could still benefit from the company’s models, thanks to Twenty. Lin’s company combines frontier AI models—though it won’t specify which — with human hacking and cyber intelligence expertise to carry out hundreds of cyber campaigns at once. Lin says Twenty does not use any of the tools to research and exploit vulnerabilities, even though models like Anthropic’s Mythos are capable of doing so.

Lin says his firm uses whatever commercially-available model suits a given task, as well as any that its U.S. military and intelligence customers already operate. Khosla Ventures partner Chu says that the company should be able to weather any loss of access to any given model. Plus, thanks to Khosla’s investment in OpenAI, Twenty can now benefit from closer ties with Sam Altman’s $852 billion AI giant.

“We have a very deep relationship with OpenAI, Vinod as the first institutional investor there, and we’re more entrenched in the West Coast,” Chu says. “The talent networks and the expertise is much deeper, so building out those relationships will be very helpful for them.”

Admiral Dennis Velez, former chief of staff at U.S. Cyber Command up until his retirement last month, expects Twenty to inspire others to join a burgeoning market supplying AI hacking tools to the Pentagon. “I think we’re just starting to scratch the surface of what's possible and what we're going to be able to do in terms of using LLMs and high-power compute to do cyber operations across the board.”