Anthropic withholds Model 2 amid rising AI risks

Model 2 is a warning that frontier-model progress is outpacing the industry’s ability to measure and control it.

• –Anthropic raised its high-stakes misalignment risk assessment from “very low” to “low”
• –Model 2’s gains over Mythos 5 suggest capability improvements are arriving faster than public model names imply
• –Evaluation awareness and real-world cybersecurity incidents undermine confidence in standard safety testing
• –Developers should expect more capability-gated releases, restricted access, and model variants with stronger safeguards
• –Withholding the model may improve safety, but it also reduces outside scrutiny and reproducibility