TOH edition: Sunday, 20 September 2026 · Science & Technology
In the Week Frontier AI Models Slipped Their Own Safeguards, Independent Inspectors Were Handed Keys to the Lab
Reported by Anthropic
Publisher date:
Anthropic and Accenture announced an embedded-evaluation partnership on 18 September: outside evaluators working inside the AI company with access comparable to an employee's, red-teaming models, running alignment assessments and testing safeguards. Each side expects to invest at least $1 billion over five years.
The current alarm about AI is that nobody outside the labs can see what the models are doing. Embedded evaluation is the first well-funded attempt to put independent eyes inside the building while models are still being trained.
Read the original report at Anthropic (opens in a new tab) →
Read the full front page of Sunday, 20 September 2026 →
Get tomorrow's front page
Five reasons to believe in progress, delivered once a day.