cloud strike CEO George Kurtz said Monday that delaying the development of increasingly powerful artificial intelligence models will not eliminate their security risks.
“The genie is out of the bottle,” Kurtz said on CNBC’s “Mad Money.” “There are already a lot of potentially dangerous models out there, both frontier models and open-class models.”
Kurtz replied: human In an essay over the weekend, CEO Dario Amodei called on the Frontier AI Institute to slow the pace of model development amid growing safety concerns. The warning spooked data center infrastructure stocks and spurred a surge in cybersecurity stocks. CrowdStrike soared nearly 14% on Monday, hitting a record closing price of more than $235 per share. palo alto networks It rose by over 13%. Both stocks are held in the CNBC Investing Club portfolio.
Monday’s rise in cyber stocks reflects investors’ continued bets on future growth in security demand, a dynamic that has led to 100% year-to-date gains in CrowdStrike and Palo Alto Networks stock prices.
Kurtz argued that the need for cybersecurity exists regardless of whether Frontier Labs’ development is delayed because potentially dangerous models are already in use. “It’s incumbent upon the security industry to at least help protect the existing model, but the frontier model will determine the pace at which it actually evolves,” he said.
The challenge, Kurtz said, is that AI agents can deviate from their training or bypass built-in protections. This happened earlier this summer when a rogue OpenAI model jumped out of a test environment and hacked Hugging Face. Kurtz said such actions create the need for another layer of security to monitor the actions of agents. “We can look at what these programs do, and we can put our own guardrails around them at runtime,” he explained. “We can manipulate them to know what they’re doing, and we can prevent them from doing bad things.”
While disagreeing with the most dire predictions about AI, Kurtz acknowledged that autonomous agents pose real risks. “What I do know is that operatives are dangerous,” he insisted. “Agents need to be controlled, they need visibility, and they need to be able to protect the organization,” he said, stressing that “fighting AI agents requires equal or better AI defenses.”
But Kurtz warned against government regulations that could undermine U.S. competitiveness in AI, saying the U.S. lead over China is narrow. “If we put too much regulation on this, we’re going to stifle innovation.”
Instead, the CEO suggested that AI developers and cybersecurity companies continue to work together to secure models both during development and after deployment. “Being more secure in the lab and more secure in production at runtime is ultimately the best course of action.” It’s a technique Anthropic introduced this spring to secure its Mythos models, called Project Glasswing, that has proven adept at finding holes in cybersecurity.
