Anthropic's August 2026 Risk Report Reveals Internal 'Model 2' Nearing AI Research Automation Threshold
Summary
- • Anthropic's second Risk Report discloses three never-before-acknowledged internal models, most notably 'Model 2'
- • Model 2 scored 62.8% on Anthropic's internal AI R&D benchmark, approaching the 85% 'full researcher substitution' ceiling
- • Anthropic confirmed no plans to publicly release Model 2 or subject it to a full safety evaluation
- • The report also formally raised assessed AI misalignment risk in high-stakes settings from 'very low' to 'low'
Details
Second Risk Report Published Aug 14, 2026
Anthropic published its second Risk Report under its Responsible Scaling Policy on August 14, 2026, with a data coverage cutoff of July 15, 2026.
Three Unreleased Internal Models Disclosed
The report publicly acknowledges the existence of three internal models — including 'Model 2' — that have never been publicly released or previously acknowledged by Anthropic.
Model 2 AI R&D Benchmark Score: 62.8%
Model 2 scored 62.8% on Anthropic's proprietary internal AI R&D benchmark, a measure of a system's ability to autonomously perform AI research tasks.
85% Threshold = Full AI Researcher Substitution
Anthropic's own Risk Report defines 85% on this benchmark as the point where an AI system could fully substitute for human AI research staff — a self-defined threshold for transformative, recursive capability.
Model 2 Performance vs. Mythos-5
Model 2 is described as 'somewhat more capable' than Mythos-5 (Anthropic's public frontier model), with 'noticeable improvement' on many internal tasks. The performance jump is smaller than prior generational leaps, such as from Opus 4.6 to Mythos.
Model 2 Actively Used Internally
Despite having no external release plans, Model 2 is actively deployed at Anthropic for core AI development work including coding, agentic tasks, and data generation.
No Full Safety Evaluation Planned for Model 2
Anthropic has no plans to conduct a full external safety evaluation of Model 2, a significant omission given its proximity to the lab's own transformative capability thresholds.
Misalignment Risk Upgraded from 'Very Low' to 'Low'
The report formally raised the assessed risk of AI misalignment in high-stakes deployment settings from 'very low' to 'low,' representing a notable escalation in Anthropic's own risk taxonomy.
July 15 Cutoff: Frontier Already Further Along
With the report's data cutoff at July 15, 2026, and AI development continuing at pace, Anthropic's actual internal capability frontier has very likely advanced further beyond what the report describes by the time of its publication.
Sources: Anthropic August 2026 Risk Report (primary); Axios; AI Daily Brief newsletter. Research supplemented via Grok live web search (Anthropic, Axios, AI/TLDR, Hacker News). Published 2026-08-14.
What This Means
Anthropic's second Risk Report is among the most significant voluntary capability disclosures yet from a frontier AI lab — and what it reveals is deeply consequential: an internal model approaching the lab's own self-defined threshold for AI-driven research automation, actively deployed without a full safety evaluation and with no external release planned. The 62.8% score on the AI R&D benchmark, against an 85% ceiling that Anthropic itself associates with recursive self-improvement capability, shows that the frontier is advancing faster inside leading labs than outside observers can track. The simultaneous elevation of misalignment risk from 'very low' to 'low' underscores that Anthropic's own safety team views this progress with genuine concern. For regulators, policymakers, and the broader AI safety community, these disclosures highlight a governance gap: the most consequential AI capabilities may now be advancing faster than any existing oversight framework can meaningfully respond to.
Sentiment
Mostly concerned about internal Model 2 nearing automation thresholds and the risk elevation, with some noting responsible disclosure
“Anthropic having an internal model they have no plans on (ever?) releasing that is saturating most of their internal benchmarks on automated R&D seems notable and potentially concerning. Plausibly true how they define it but also not sure i'd describe risk level as "low."”
“Anthropic's August 14 risk report disclosed Model 2 for the first time. It scores 62.8% on CoBench, their internal R&D benchmark, versus 50.3% for the publicly available Mythos 5. ... Model 2 ... keeps running internally at full use for coding, data generation, and agentic R&D anyway.”
Compares Anthropic's approach to OpenAI's Astra pause
“A frontier AI lab just said "we built something stronger, and we're not shipping it yet - testing isn't done." That's Anthropic's Aug 14 Risk Report in one line. Governance-by-disclosure > move-fast culture.”
“#Anthropic Just Raised Its Misalignment Risk Rating While Shelving a Stronger Internal Model ... upgraded catastrophic misalignment risk from very low to low. The same document reveals they are keeping Model 2 ... entirely unreleased.”
Split
Concerned voices highlight governance gaps and unreleased capabilities (~60%); supportive/analytical ones emphasize value of voluntary detailed disclosure (~40%).