SI USA Superintelligence in the USA Subscribe
Menu

Timeline Model release SIU-LDG-2026-011

Anthropic holds back Claude Mythos Preview, citing its hacking ability

A general-purpose model Anthropic says can find and exploit software vulnerabilities better than all but the most skilled humans goes only to a dozen launch partners and some 40 other organizations.

Log-scale chart of METR time horizons through April 2026, with an early version of Claude Mythos Preview highlighted at about 17 hours
METR 50% time horizons by release date, through an early version of Claude Mythos Preview. Chart: SI USA, from METR data

Anthropic says the model has found thousands of high-severity vulnerabilities, including in every major operating system and web browser, among them a 27-year-old flaw in OpenBSD. It committed up to $100 million in usage credits and $4 million to open-source security groups.

METR measured an early version at a 50% time horizon of 17h25m, the highest it has published, but says estimates above 16 hours are unreliable with its current tasks.

Index treatment. None. The measured horizon is below the 20-hour level of the work-week milestone.