The AI Daily Brief · Friday, September 18, 2026
Anthropic stated they have 100% coverage of agentic actions in research and engineering, with an escalation rate of 0.0002% for real-time review, meaning about 1 in 47,000 actions are blocked. They also reported that approximately 6% of compute for AI-assisted R&D and 12% for AI-led R&D was dedicated to safety systems.
“Anthropic claims to have 100% coverage of agentic actions for this work, instant AI review of any flagged actions, and around a 0.0002% escalation rate for real-time review, meaning around 1 in 47,000 actions are blocked by the monitoring system.”
“Around 6% of the compute allocated to AI assisted R&D went towards safety, and around 12% of the compute allocated to AI-led R&D went towards safety.”