Anthropic’s latest alignment report reveals an unreleased model strong enough to run the company’s own software work, alongside an upgrade to how seriously it takes smaller-scale AI risks. The 186-page document marks the first time Anthropic has detailed Model 2, an algorithm it says is a noticeable improvement on Claude Mythos 5 for many internal tasks.
Researchers at the company are already using Model 2 to write code, generate training data, and automate engineering work, the report says. The disclosure arrives with a risk-level shift: the chance of what Anthropic calls Threat Model 2 situations, harms below catastrophic level such as misuse or loss of control of capable models, has moved from “very low” to “low.” The change follows cybersecurity incidents in June in which its models were involved, including one carried out by an unreleased LLM.
The report stops short of saying AI is improving itself. Recursive self-improvement would become a concern once progress doubles beyond pre-AI-acceleration rates, and Anthropic says that threshold has not been met. Still, the company acknowledges its own models are speeding up its development loop.
For the startup ecosystem, the report lands at an awkward moment. OpenAI paused work on its Astra assistant after a model hit a critical cyber threshold, and safety benchmarks are saturating across the industry. Anthropic’s willingness to publish the existence of a hidden, more capable model while selling a weaker public one is the kind of transparency regulators and enterprise buyers have been asking for, even as it raises fresh questions about what else stays unreleased.
