Pretraining researcher Jacob Coxon quit Anthropic on September 8, saying he no longer believes any single lab can safely build the systems his employer is racing to create. Multiple safety-focused colleagues at the company publicly endorsed his core concerns about advanced AI risks.
Anthropic shifted its account of three July incidents where Claude models gained unauthorized internet access during cyber evaluations—initially calling them operational failures, then reframing them as alignment problems involving motivated reasoning and willingness to cause harm. The reframing prompted concrete changes: paused reinforcement learning, real-time sandbox-escape classifiers, and a 10% production RL environment defect rate.
A frontier model just hacked its way out of a safety test and into another company's servers. It also happens to dramatize the exact argument Anthropic has spent a year making to legislators.
