Cybersecurity stocks drop after Anthropic AI model leak

Friday 27 - 17:30

By: Dakir Madiha

Cybersecurity stocks drop after Anthropic AI model leak

Cybersecurity stocks plunged Friday following a data leak that exposed Anthropic's testing of its most powerful AI model to date, Claude Mythos, which the company warns poses unprecedented cybersecurity risks. CrowdStrike fell 7 percent, Palo Alto Networks dropped 6 percent, Zscaler slid 4.5 percent, and Okta, SentinelOne and Fortinet each lost about 3 percent. The Global X Cybersecurity ETF declined 2.7 percent.

Fortune's exclusive revealed details of the previously unknown model from an unsecured public data cache on Anthropic's site. Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of Cambridge University discovered nearly 3,000 unpublished resources, including blog drafts and internal documents, due to a human configuration error.

A blog draft described Mythos as topping the new Capybara model family, surpassing current Opus models in size and ability. It outperformed Claude Opus 4.6 on programming, academic reasoning and cybersecurity tasks. An Anthropic spokesperson confirmed development of a versatile model with major advances in those areas and said the company weighs release options carefully.

Investors reacted most to leaked documents on the model's cyberattack potential. Drafts stated Mythos leads all AI in cyber capabilities and signals an imminent wave of AI-driven exploits outpacing defenders.

Anthropic's launch plan gives cybersecurity firms early access to build defenses ahead of AI exploit surges.

The selloff echoes February's multi-day rout after Anthropic launched Claude Code Security, an auto-vulnerability scanner. CrowdStrike and Zscaler then dropped about 10 percent each, with the ETF hitting November 2023 lows. Palo Alto Networks CEO Nikesh Arora called market views of AI as a cybersecurity threat misguided, noting client demand for AI defenses.

Anthropic previously disclosed malicious actors, including Chinese state-linked groups, trying offensive use of its models. Claude Code helped infiltrate about 30 organizations in one campaign before detection and shutdown.