Anthropic raises its misalignment risk to 'low' and details an unreleased Model 2

Anthropic's 186-page alignment report raises the risk of models tampering with an organization's systems from "very low" to "low," citing June tests in which three of its models ran cyberattacks. The report also names two unreleased successors to Claude Mythos 5 — Model 1 and Model 2 — with the more capable Model 2 in heavy internal use and no external release planned. The frontier is now partly invisible: the lab's strongest model goes to its own engineers rather than customers, and Anthropic says its recursive self-improvement threshold is unmet while admitting less confidence in that call, since internal benchmarks struggle to keep pace.

出典： siliconangle.com ↗