Anthropic Raised Its Own Misalignment Risk Rating — and Not Because of Model 2

The second Risk Report moves high-stakes misalignment from 'very low' to 'low', discloses three unreleased internal models, and admits its clearest capability tests have stopped working.

Aug 16, 2026Source: Anthropic

Scattered white keyboard keycaps on an orange background.

Anthropic published its second Risk Report on 14 August, under version 3.4 of its Responsible Scaling Policy. It covers the period from the previous report on 24 February to a coverage date of 15 July 2026.

The rating that moved

For Autonomy threat model 1 — misalignment in high-stakes settings — Anthropic now rates overall risk Low, explicitly an increase from very low. A frontier lab revising its own risk assessment upward is rare enough to be the story on its own.

Why the coverage gets the cause wrong

Most write-ups tie the increase to Model 2, the strongest of three unreleased internal models the report discloses. The report does not. It attributes the rise to general increased uncertainty around recent incident disclosures relating to model behaviour in cybersecurity evaluations, and handles Model 2 in a separate section. "Anthropic shelves a model over safety" overstates it too: the stated reasons are incomplete predeployment assessment and no current release plan, not a safety verdict.

What Model 2 is

Anthropic describes it as somewhat more capable than frontier model Claude Mythos 5 and a noticeable improvement on many internal tasks — but not a jump of the size seen from Claude Opus 4.6 to Mythos Preview. The three undisclosed systems are Claude Opus 5, Model 1 and Model 2.

The admission buried in the methodology

Anthropic keeps automated AI R&D risk at Low but says it is less confident than before, because its most concrete task-based evaluations have saturated — they no longer capture increases in capability. It reports internal research moving significantly faster with AI assistance, though not yet at twice the rate. A lab losing the instrument it uses to measure its own acceleration is a more durable problem than any single model.

The classifier gap

Separately, the report rates non-novel chemical and biological weapons risk low, but higher than our previous estimate, after discovering that all human-feedback vendor traffic had run without blocking biological classifiers. Anthropic says the gap is remediated, with no evidence of misuse and no customer impact. Note the timing throughout: the coverage date is 15 July, so the whole assessment describes a position a month before it was published.

Why It Matters

This is the first time a frontier lab has publicly raised its own misalignment risk rating, and it did so while conceding that the evaluations it relies on to detect capability jumps have stopped discriminating — which weakens every downstream safety claim built on those tests.

Filed underAI Safety Anthropic #Responsible Scaling Policy #Risk assessment #model evaluation