Armadin's Kevin Mandia says AI agents need cyber experts

Kevin Mandia, who founded and runs the cybersecurity company Armadin, said securing AI that works on its own takes real security expertise. He said so on the a16z Podcast, in an episode posted Tuesday, Oct. 6.

This matters to anyone whose bank, hospital or employer relies on software that AI is starting to run. If AI tools get hacked or go wrong, the cost lands on customers. Mandia thinks AI teams and security teams have to work side by side.

He said Armadin built AI that attacks company networks, and it breaks in well. To keep it in check, the team watches every instruction it gets and logs each action. They also set firm rules that block certain moves. In a test of 20 real attacks that humans had carried out, no AI got past eight. Mandia said open models, which anyone can download and run for free, matched the best closed ones. He put it in a short line at 30:17.

"we tested the open weight ones and the most advanced closed models, they all found eight."

The closed models were faster, he said, but the open ones got to the same place when given more time.

CNBC on MSN published a story on Oct. 19 headlined "Armadin CEO Kevin Mandia on defending against AI cyber threats, future of AI agents." It describes him joining Squawk Box to talk about how businesses and boards are adapting.

Mandia also said too many limits hold back the AI's creativity, so there is a gray area. When the team can't tell what the AI is doing, a human decides. Armadin builds the AI attack tools it describes here.

Watch the 8 minutes where he explains how Armadin keeps its AI in check.27:02