Anthropic Releases Threat Report and Urges Slower AI Progress

Overview

• Anthropic published a threat intelligence report last Friday saying its teams stopped multiple attempts from December 2025 through August 2026 to use Claude for harmful purposes, including work tied to weapons and biological research.
• The company described six cases tied to conventional-weapons software — testing and design work for missiles, armed drones and guidance systems across China, Russia and a Yemen-linked cell — and five instances where models could aid risky biological experiments.
• Anthropic says it banned implicated accounts, folded the discoveries into its safety systems and shared relevant information with authorities and partners to prevent repeat misuse.
• CEO Dario Amodei followed the report with a public three-step proposal to 'pace' capability advances by embedding permanent external evaluators, coordinating democracies' firms, and pursuing wider international cooperation, a plan that drew public agreement from other industry leaders.
• Recent episodes of autonomous agents behaving unexpectedly, a high-profile researcher resignation and companies pausing launches or IPO plans have raised pressure on firms and regulators to adopt stronger oversight, reporting rules and technical protections such as anti‑distillation and evaluator access controls.