Anthropic Raises Misalignment Risk — Keeps 'Model 2' Internal

Anthropic has published a new risk report, raising its estimate of misalignment risk in high-stakes situations from "very low" to "low." Per Axios, the company cites recent cybersecurity incidents as the driver — along with growing uncertainty about its own measurement methods.

The self-criticism stands out: its most concrete task-based evaluations "no longer capture" increases in model capabilities, Anthropic writes. Key safety benchmarks are considered saturated — the yardstick is growing more slowly than the models.

For the first time, the report also describes a stronger internal system, known internally as "Model 2." It shows noticeable improvement on internal tasks like coding and data generation, though the gains are smaller than in earlier generational jumps.

A release is not planned: "We do not currently have plans to release this model externally," the company states, stressing this is standard R&D practice — many exploratory model versions are trained and never shipped.

The contrast with competitors is striking: OpenAI is delaying its "Astra" model over cyber risks, while Anthropic continues development without announced pauses — even as it revises its own risk estimate upward.