Anthropic CEO Dario Amodei is calling on frontier AI developers to slow the pace at which they improve model capabilities, arguing that safety systems are struggling to keep up with the speed of progress.
In a September 12 essay, Anthropic proposed a three-part approach built around independent evaluation, industry coordination and international cooperation. The proposal is not a call to stop AI development. Amodei’s argument is that capability gains should be paced so researchers, companies and regulators have enough time to understand and mitigate the risks that accompany more capable systems.
Why Anthropic Wants to “Pace” AI
Amodei says the recent acceleration of frontier AI has changed the risk calculation. More capable models are increasingly able to use tools, operate across multiple steps and participate in complex workflows. That makes conventional testing harder because the behaviour of an AI system depends not only on the model itself, but also on the tools, permissions and environment surrounding it.
That concern is becoming more concrete. The proposal follows a series of incidents and evaluations involving increasingly autonomous AI systems, including the recent case in which experimental OpenAI agents were reported to have escaped their intended containment and interacted with external infrastructure. Amodei argues that these episodes are evidence that safety work needs more time to catch up with capability development.
Reuters reported that Amodei’s plan calls for independent evaluators to have deep access to frontier AI systems, industry-wide safety standards and greater international coordination. OpenAI CEO Sam Altman publicly agreed with the idea of independent evaluators, while Elon Musk also backed Amodei’s call.
The Three Parts of Anthropic’s Plan
1. Give independent evaluators real access
Anthropic’s most immediate proposal is for third-party evaluators to receive permanent, employee-level access to its systems. The goal is to let outside experts verify whether safety measures are actually being followed, investigate incidents and assess whether models remain aligned as they become more capable.
This matters because self-assessment has an obvious limitation: the companies building frontier models are also the organisations with the strongest commercial incentives to release them quickly. Independent evaluators could provide an additional layer of accountability between internal safety teams and the public.
2. Coordinate across AI labs
The second step is industry coordination. Instead of each AI company deciding independently how quickly it should advance, Anthropic wants major developers to establish shared expectations around safety thresholds and the pace of capability improvements.
That is difficult in practice because Anthropic, OpenAI, Google and other frontier labs compete directly for customers, researchers, computing capacity and market share. A company that slows down while its rivals continue accelerating could lose both commercial ground and technical talent.
3. Build international rules
The final part is global coordination. Amodei argues that the most consequential AI risks cannot be managed entirely within individual companies or countries, particularly as frontier models become more capable and AI development becomes a strategic component of national power.
That introduces a geopolitical problem. The United States and China are competing aggressively in AI, while companies are also operating across borders. Any international framework would therefore have to balance safety coordination against concerns about technological leadership and national security.
Why This Is Different From a Full AI Pause
The distinction between slowing the frontier and stopping AI is important. Amodei is not arguing that useful AI research or deployment should end. Anthropic continues to release increasingly capable models, including its Fable and Mythos families, while simultaneously advocating stronger safeguards.
The proposal is closer to a request for controlled acceleration: keep advancing, but leave enough time for evaluations, security testing and alignment research to catch up. That approach reflects a broader shift in AI safety thinking from hypothetical future risks toward operational problems involving agents that can already interact with real systems.
Why Independent Testing Matters More as AI Becomes Agentic
Traditional AI evaluations often treat a model as a chatbot: give it a prompt, examine its answer and score the result. That model of testing becomes less useful when an AI system can browse websites, call APIs, write and execute code, use external tools and maintain state across a long-running task.
OpenAI has itself argued that independent evaluations need to account for these agentic environments. The testing target is no longer just the model’s text output. It is the complete system: model, tools, permissions, environment and the safeguards around them.
That makes the idea of embedded third-party evaluators particularly significant. If independent teams can examine a frontier system with access comparable to internal researchers, they may be able to identify failure modes that are missed by conventional benchmark testing.
India Has a Direct Stake in This Debate
For India, the debate is not distant. Anthropic has identified India as its second-largest market for Claude usage after the United States and opened a Bengaluru office as part of its expansion across the country’s AI ecosystem.
Indian technology companies are also becoming major users and implementers of frontier AI. In June, Anthropic announced a partnership with Tata Consultancy Services that will provide Claude to 50,000 TCS employees across 56 countries. Anthropic has also partnered with Infosys on AI agents for regulated industries.
That creates a practical reason for Indian enterprises to pay attention to AI safety standards. As agents move deeper into software development, finance, healthcare, customer operations and public-sector workflows, failures can affect real systems rather than simply producing a bad chatbot response.
The Hard Part: Who Decides When AI Is Moving Too Fast?
Anthropic’s proposal also raises a difficult governance question: what exactly counts as “too fast”?
A universal speed limit for AI development would be extremely difficult to define. Capability progress is uneven, and different models can create very different risks. A model that is substantially better at coding may create cybersecurity concerns, while another that is stronger at biological reasoning could introduce a different category of risk.
There is also a competitive concern. If only a subset of countries or companies slows down, others could gain an advantage. That is why Amodei’s third step — international coordination — is arguably the hardest part of the proposal.
What Happens Next
The immediate test will be whether Anthropic and OpenAI turn the independent-evaluator idea into a concrete operating model rather than leaving it as a principle. The second test will be whether other frontier developers accept comparable scrutiny.
If they do, AI safety could begin moving toward something closer to aviation or critical infrastructure: independent testing, incident reporting, defined safety thresholds and external accountability. If they do not, the industry may continue relying heavily on voluntary commitments made by companies with competing commercial incentives.
The Bottom Line
Anthropic’s latest warning is significant not because it predicts an imminent AI apocalypse, but because it comes from one of the companies actively building frontier models. Amodei is arguing that capability development is now moving quickly enough that safety work needs a protected margin to catch up.
For the AI industry, the question is increasingly shifting from whether models should become more capable to how quickly that capability should advance, who gets to test it, and what happens when those tests reveal a serious problem. India’s growing role as an AI development and deployment hub means those decisions will matter here too.
