According to reporting by The Decoder, the nonprofit METR has concluded an eleven-week audit of Anthropic’s internal research data. The findings indicate that the pace of model development at the company roughly doubled, with AI-assisted research driving about 80 percent of this speedup. METR asserts that this acceleration may have crossed Anthropic’s own internal warning threshold as early as July.
Audit Limitations and Redactions
During the assessment, Anthropic redacted the most informative evidence from the data provided to METR. Consequently, the public must rely on METR’s independent judgment to a degree that METR describes as much more than would be ideal. In addition, skeptical voices reportedly exist within Anthropic regarding these developments, though the exact nature and scope of internal dissent were not detailed in the available information.
What Remains Unclear
Significant gaps remain regarding the specific internal metrics Anthropic uses to define its safety thresholds. The announcement did not specify the exact criteria that triggered the warning level, nor did it disclose which internal projects or teams experienced the most significant speedups. Observers will need to watch for future disclosures from either organization to understand the long-term safety implications of AI systems being used to accelerate their own successors.
As the intersection of AI safety research and recursive self-improvement develops, the transparency of independent audits will remain a central point of discussion.
Lead image: Matheus Bertelli / Pexels
