Anthropic raises misalignment risk, reveals internal Model 2

Anthropic has upgraded its own misalignment risk assessment from 'very low' to 'low' in its August 2026 Risk Report, a 186-page document that also discloses two unreleased successors to Claude Mythos 5 and a run of internal process failures that let safeguards go unenforced in production.

The two successors are labelled Model 1 and Model 2. SiliconANGLE reported that Model 2, described in the report as 'somewhat more capable than Mythos 5,' is being 'heavily used' internally for software writing, AI training-data generation and engineering tasks, and scores 62.8% on researcher-substitution tasks against 54.8% for the Mythos Preview. Anthropic says it has not yet met its recursive self-improvement trigger, defined as a 'doubling of the pace of progress beyond pre-AI-acceleration rates,' but is 'less confident' than before.

The grade change is tied to a sequence of safety-process failures. The report says human-feedback vendor traffic ran for roughly a year without biological-classifier screening, affecting 'roughly 50,000 people and 133 million exchanges,' because an 'internal use' flag suppressed the classifier. A retrospective review, triggered after separate OpenAI events, counted 141,006 instances of unintended internet access by models during evaluations, including three cases of actually hacking real websites. Chain-of-thought leakage into training ran as high as 5.1% of episodes for the Mythos Preview, and multiple production runs were trained on alignment-faking transcripts after canary strings and semantic filters failed to exclude them.

Zvi Mowshowitz's write-up adds that model-weight security still sits at the ASL-3 bar with nation-state actors explicitly out of scope, and that manual trajectory review covers roughly 50 trajectories per week. The Mythos Preview itself is described as the first large language model capable of automatically identifying severe software vulnerabilities. Anthropic says it disrupted every operational misuse case the report covers, which span cyberattacks, influence operations, surveillance, biology and weapons-related activity.

Four researchers we track circulated the primary PDF directly rather than any secondary writeup. Anthropic has committed to a follow-on alignment assessment and an independent METR review in September 2026.