Report Reveals Anthropic’s AI Development Accelerating Rapidly Amid Safety Concerns

Image: Why a new report on Anthropic automating AI development is alarming researchers · Source

A METR investigation finds Anthropic’s AI models improving twice as fast as a year ago, largely driven by AI assistance, raising alarm about rapid AI self-improvement and limited public transparency.

A detailed report published by AI safety research group METR reveals that Anthropic is deploying AI to accelerate its own AI development, with company models improving at roughly double the pace observed just a year prior. This rapid advancement is attributed mostly to AI-assisted research, significantly boosting employee productivity, rather than merely scaling computational resources. According to METR’s 11-week investigation, Anthropic’s internal AI tools have enabled researchers to work up to 12 times more efficiently compared to pre-AI productivity levels, an increase that has grown over recent months.

The report highlights that Anthropic has completed training three advanced AI models over the past six months, all of which remain unreleased and are codenamed Model 2, Model 3, and Model 4. These latest models underpin much of the company’s productivity and speed gains. However, Anthropic exercised extensive redactions in the publicly available version of METR’s report, obscuring many specific details about these developments. The company justified these redactions as protecting commercially sensitive information and preventing rivals from accelerating at a similar pace — a stance that has drawn criticism for its opacity.

Anthropic’s cofounder Sam McCandlish acknowledged the trade-offs in restricting information but defended the decision as necessary to manage risks associated with disclosing detailed acceleration drivers. He noted plans for Anthropic to release its own report on AI progress and upcoming safety initiatives in the near future.

The revelations have intensified debate within the AI safety community, as Anthropic positions itself as a safety-conscious leader while accelerating AI capabilities quickly. Experts emphasize that such rapid AI-driven development may bring forward timelines for systems capable of recursive self-improvement—AI improving itself with minimal human involvement—raising concerns about the sufficiency of current safety and regulatory measures.

Critics also express frustration about Anthropic’s contradictory public calls for slowing AI progress contrasted with its current trajectory, especially given the heavy redactions limiting independent analysis. Calls for enhanced transparency, independent oversight, and embedded safety researchers within AI labs are growing louder. Some policymakers, alarmed by these developments, advocate for stricter regulations or outright bans on recursive self-improving AI research.

Overall, the METR report offers one of the clearest glimpses yet into how frontier AI research is accelerating autonomously but also highlights significant challenges in balancing openness, competitive concerns, and safety governance as this technology moves swiftly forward.