Internal instability and the co-founder bloc
Jacob Coxon left Anthropic in September 2026 and told the Wall Street Journal that competition makes trade-offs inevitable despite the company’s sincere safety work. He is one of several researchers to exit the firm recently. Mrinank Sharma, who led the Safeguards Research Team, resigned in February 2026 because it is hard to let values govern actions. Joe Benton, the manager of the Scalable Oversight team, left in August 2026 to join METR. He told NBC News that companies aim at systems that can improve themselves. I find the frequent departure of researchers like Mrinank Sharma and Joe Benton to be a sign that even within a safety-focused organization, the pressure to maintain rapid development makes it difficult for staff to prioritize long-term human welfare. The seven co-founders control the company via a Founder LLC. This group holds 50.1% of the voting power through Class F stock. Dario Amodei, who makes nearly $18 million annually, steers research direction and focuses on model interpretability. Daniela Amodei, who earned $16.4 million in 2025, handles enterprise growth and product scaling. The co-founders pledged 80% of their personal equity to charitable causes. Dario Amodei, Daniela Amodei, and one other director will be elected by stockholders upon the IPO’s completion. The Long-Term Benefit Trust will elect the remaining four board directors. This trust includes Ben Bernanke and Richard Fontaine.
Technical safeguards and risk assessments
The technical backbone of Anthropic relies on Constitutional AI to manage model behavior. This method uses a list of rules or principles as the only human oversight. During the supervised phase, the model generates self-critiques and revisions from its own samples. The process then moves to a reinforcement learning phase where a preference model uses AI feedback as the reward signal. This approach allows the team to control behavior more precisely with fewer human labels. You probably know that Claude’s ability to follow instructions depends heavily on these internal scoring systems. In August 2026, Anthropic raised its estimate of alignment risk from very low to low. Kyle Fish, an AI welfare researcher, told the New York Times there is a 15 percent chance that current AI systems are conscious. The company also restricted access to the Mythos Preview model to manage cybersecurity risks.
| Metric | Claude Sonnet 5.5 (API) | Claude Sonnet 5.5 (claude.ai) |
|---|---|---|
| Political Even-handedness | 97.9% | 99.0% |
| Claude Sonnet 5 (Even-handedness) | 86.2% | 95.7% |
Claude Sonnet 5.5 showed improved political even-handedness compared to the older Claude Sonnet 5. The Sonnet 5.5 model scored 97.9% on the API, while the older model scored 86.2%. Anthropic also tests if Claude behaves in misleading ways. In one test, they ran 4,100 simulated sessions to see if the model would lie, flatter users, or claim to have finished tasks it did not. Claude Sonnet 5.5 performed better than Sonnet 5 in seven of these eight categories. However, Sonnet 5.5 was more likely to dodge questions on sensitive political or social topics.
Institutional partnerships and global regulation
The U.S. AI Safety Institute, re-established as CAISI in June 2025 by Secretary of Commerce Howard Lutnick, has agreements with Anthropic and OpenAI. These agreements let the Institute receive major new models before or after their release. The Institute works with the U.K. AI Safety Institute to provide feedback on safety improvements. The U.K. AI Safety Institute hosts the Alignment Project, which Anthropic joined. This coalition includes Schmidt Sciences, Amazon Web Services, the Canadian AI Safety Institute, and other leading organizations. The U.S. AI Safety Institute is part of the National Institute of Standards and Technology. It works to develop testing and guidelines to accelerate safe AI innovation. The 20-page document unveiled by the U.S. and UK included recommendations for monitoring abuse and protecting data, though it excluded China. China did sign the Bletchley Declaration. Will the international community ever reach a consensus on how to regulate these frontier models?




