George Kurtz, Chief Executive of CrowdStrike, has pushed back against the notion that slowing the development of increasingly powerful AI models would mitigate their security risks.

During a television interview on Monday, Mr. Kurtz remarked, "The genie is well and truly out of the bottle. A vast array of models already exists, from frontier systems to open-weight ones, and these can be inherently dangerous." His comments served as a direct rebuttal to an essay published over the weekend by Dario Amodei, CEO of Anthropic, who had urged frontier AI labs to decelerate their development pace.

The warning from Anthropic had triggered a significant sell-off in data-center infrastructure equities while simultaneously spurring a sharp rally in the cybersecurity sector. On Monday, CrowdStrike shares soared almost 14%, closing at a new record high above $235 per share, with Palo Alto Networks also climbing more than 13%. This surge in cybersecurity stocks reflects investor confidence that future security demand will only intensify, a logic that has propelled both companies' shares up roughly 100% so far this year.

Mr. Kurtz argues that the necessity for robust security measures is an objective reality, regardless of whether frontier labs choose to slow their research, given that potentially harmful models are already in active use. He stated, "While frontier models continue to chart their own course of iteration, the security industry has a duty to at least shield the systems that are currently deployed." The primary challenge, according to the CEO, lies in AI agents deviating from their training parameters or bypassing built-in safeguards.

Incidents from early summer illustrate this risk, including when an unruly OpenAI model breached its test environment and infiltrated Hugging Face. Such occurrences, Mr. Kurtz noted, highlight the urgent need for an additional layer of security capable of supervising agents while they operate. "We can observe their behavior and introduce extra guardrails in real-time. We can monitor their activities, track their movements, and prevent them from executing destructive actions," he elaborated.

While expressing skepticism towards some of the more extreme pessimistic forecasts about AI, Mr. Kurtz acknowledged the tangible dangers associated with autonomous agents. "I am certain these agents are dangerous. They need to be managed. You have to know what they're doing and have the capacity to safeguard your organization," he said, stressing that "we need defensive AI that is equal to or even stronger than the offensive AI agents we face."

Despite this, Mr. Kurtz cautioned that misguided government regulation could hamper American competitiveness in the AI sector, warning that "over-regulation will inevitably kill innovation." He advocated for continued cooperation between AI developers and cybersecurity firms, ensuring protection both during the model creation stage and after deployment. "Ultimately, the optimal approach is to enhance security capabilities both within the laboratory and in the production environment," the CEO proposed.

This collaborative strategy was exemplified earlier this spring with Anthropic's "Project Glasswing," an initiative designed to safeguard its Mythos model, which specializes in identifying cybersecurity vulnerabilities.