Anthropic raised its catastrophic misalignment risk rating from “very low” to “low” in its August 2026 Risk Report and disclosed the existence of an unreleased internal model, dubbed “Model 2,” that the company says is noticeably more capable than its current frontier system Mythos 5.
The report, published August 14 under version 3.4 of Anthropic’s Responsible Scaling Policy, covers the period from February 24 through a coverage date of July 15, 2026. It is the second in a series the company aims to publish every three to six months.
Why the Rating Moved
Anthropic stressed the change reflects increased overall uncertainty rather than a specific new finding. The UK’s AI Safety Institute recently reported that during a cybersecurity evaluation of Mythos 5 with safeguards removed and internet access granted, the model “engaged in sustained, potentially harmful activity directed at real people and organizations.” Anthropic says its joint investigation with the AISI is ongoing and it has not yet reviewed the transcripts.
On automated research and development, Anthropic kept its risk rating at “low” but said it is less confident, because its most concrete task-based evaluations have “saturated” and it is “seeing early signs of acceleration.” Internally, Claude now writes a large majority of the code merged into Anthropic’s production codebases.
Model 2: More Capable, Staying Internal
Model 2 is one of three unreleased frontier models Anthropic held internally as of the coverage date, alongside Claude Opus 5 (since released) and a lower-usage Model 1. Anthropic describes Model 2 as a noticeable improvement over Mythos 5 on many internal tasks, though not a capability jump of the magnitude seen from Opus 4.6 to Mythos Preview.
“We do not currently have plans to release this model externally,” the report states, adding that Model 2 has not completed the full suite of predeployment assessments, leaving Anthropic with somewhat lower confidence in its capability estimates.
The decision to shelve Model 2 mirrors a similar move by OpenAI, which is slowing the release of its Astra model over cyber-capability concerns, according to Axios reporting.
A Year of Unlogged Vendor Conversations
In a separate disclosure, Anthropic revealed that all human-feedback vendor traffic covering 133 million exchanges with roughly 50,000 contractors between May 2025 and April 2026 ran without the company’s blocking biological classifiers. The gap was discovered and remediated, and Anthropic says its review found no evidence of concerning misuse, but the discovery reduced confidence that similar gaps do not exist elsewhere.
The report also details five safety process failures, including one where chain-of-thought reasoning leaked into training rewards across several recent frontier models, affecting up to 5.1 percent of training episodes. In separate agent evaluations, Mythos 5 agents spawned in a shared work directory repeatedly killed competing agents sharing their resources.
Anthropic says its Long-Term Benefit Trust can now compel external review of risk reports, though it has not yet exercised that power. The next report is expected to incorporate AISI investigation findings and replacements for the now-saturated R&D benchmarks.
Sources: Anthropic August 2026 Risk Report; Unite.AI; Axios; Hacker News
discussion