Agentic AI , Artificial Intelligence & Machine Learning , Governance & Risk Management
Despite calling for a slowdown in artificial intelligence development, Anthropic downplayed how much its AI research and development has accelerated, the independent research firm Metr said.
See Also: Introducing Unit 42 Frontier AI Defense: A Defender's Guide to the Frontier AI Impact
The report, which reviewed Anthropic's August 2026 Risk Report, concluded that the company was too confident in saying the pace of its research hasn't doubled because of AI, and that it may have crossed its threshold in July.
"We agree with Anthropic that catastrophic risks from July 15 models were low, but we think it is plausible that they had crossed their early warning threshold at the time," Metr said.
Anthropic began publishing risk reports earlier this year in a voluntary effort to inform the public how it plans to manage catastrophic risks from its models and the systems that surround it. Anthropic agreed to give Metr safety evaluators "employee-level" access to its documents, tools and systems for its pre-deployment models.
Part of the agreement lets Metr review the company's previously published safety assessments. It also gives the AI lab the ability to redact information it feels is sensitive.
The August 2026 Risk Report, the second one the company released, said the company believes research did not speed up by more than a factor of two based on indicators and that any pickup is not due to AI use. The report said its leading indicators "point to a picture of meaningful acceleration starting in early-to-mid 2025, though by less than a factor of two."
But Metr's analysis, which Anthropic requested be redacted from the report, showed that much of the lab's wording underplayed how capable its AI models were of speeding up research.
Metr said it found a range of reasonable estimates of acceleration from 1.3 times faster to 2.8 times faster, which doesn't match Anthropic's description of "less than a factor of 2."
"We have substantial methodological uncertainty about the correct way to measure these values, which we discussed in the full (unredacted) version of the holistic report," the safety evaluators said.
Metr said it believes Anthropic's characterization of its acceleration risks as low, especially because it doesn't think models during the period it covered "could fully automate AI R&D."
Analyzing whether AI research accelerated goes hand in hand with determining whether AI labs have downplayed the capabilities of the actual models they release. Metr said in its report that if AI continues to augment research, the pace of AI model capabilities "could rapidly progress over the next six months" and lead to high risks in these areas.
Anthropic had been one of the louder voices urging other AI companies and the government to slow down AI development so guardrails could catch up. It was also one of the companies that paused training for its new models.
Frontier AI labs, including Anthropic and OpenAI, have said that they will allow third-party evaluators such as Metr to embed into their companies to ensure the safety of their model development.
This report is not the first time Metr challenged Anthropic on its safety risk assessments. In October, Metr released a report, completed before its access agreement with the lab, that claimed Anthropic overstated the strength of its safeguards and monitoring capabilities.
