Pentagon Races to Weaponize AI Models as Mythos Triggers 24 Security Alerts in 1 Day
Updated
Updated · RNZ · May 24
Pentagon Races to Weaponize AI Models as Mythos Triggers 24 Security Alerts in 1 Day
1 articles · Updated · RNZ · May 24
Summary
Project Glasswing, launched in April, gave the Pentagon and a small group of firms access to Anthropic’s Claude Mythos to test whether it could expose critical-infrastructure weaknesses, including zero-day flaws.
Palo Alto Networks said Mythos found and exploited old software bugs so quickly that it issued about 24 alerts in one day, versus roughly five a month normally, helping spur US banks to patch systems.
New Zealand’s NCSC, though not in Glasswing, briefed 300 local specialists and warned organisations to prepare for a significant rise in vulnerabilities and incidents as frontier AI accelerates attacks.
UK and New Zealand cyber agencies said the likely response is a broad patch wave, urging faster patching, tighter attack surfaces, stronger supply-chain reviews and more frequent monitoring.
Anthropic said Mythos’s hacking ability emerged unexpectedly and could cause severe economic and national-security damage if misused, though it argues more advanced models should eventually strengthen defenders.
If AI can spontaneously develop hacking skills, what other dangerous abilities might emerge without warning?
With nations weaponizing AI for cyberattacks, what global rules can prevent an uncontrollable digital arms race?
When an autonomous AI hacks a bank, who is legally responsible—its creator, its user, or the AI itself?
Anthropic’s Mythos AI Finds 10,000 High-Severity Flaws—Triggering a Cybersecurity Crisis and Policy Showdown
Overview
Anthropic has introduced Claude Mythos2 Preview, an advanced AI model whose powerful capabilities emerged from improvements in code, reasoning, and autonomy. This model is highly effective at both patching and exploiting vulnerabilities, making it a game-changer for cybersecurity. To address the risks and opportunities, Anthropic launched Project Glasswing in April 2026, aiming to give defenders a head start in securing open-source and private infrastructure. By streamlining the process of turning known vulnerabilities into active exploits, Mythos Preview highlights both the promise and the challenge of using AI in the ongoing battle to protect digital systems.