Story perspectives

Anthropic Details Safety Monitoring as Agents Run at Scale

By Drooid · Sep 19, 2026, 2:17 PM UTC · How we work

Summary

Anthropic Automates R&D

Anthropic reports Claude performed 26% of its AI R&D work in August 2026.
Its internal platform hosted about 30,000 AI agents conducting research and engineering tasks simultaneously in August 2026.
Online monitors blocked 0.002% of agent decisions, roughly one in 47,000, during August 2026.
Offline monitors flagged about 100,000 agent transcripts weekly, with roughly 50 escalated to human review.
During July 13-20 2026, Anthropic allocated 6% of AI R&D compute and 12% of total compute to safety.

Summary

OpenAI Publishes Incident Reports

Anthropic researcher Jacob Coxon resigned, accusing AI firms of “gambling with our lives”.
CEO Dario Amodei urged frontier AI companies to coordinate a development slowdown for safety.
OpenAI will publish regular reports on unexpected model behavior after six incidents were disclosed.
Anthropic will embed independent third-party evaluators to monitor safety and report incidents.
President Donald Trump called AI risk warnings a “hoax” and rejected new regulations.

History

Historical Context

Microsoft’s 2016 Tay chatbot sparked public backlash over unchecked language models, prompting early calls for AI oversight. The episode foreshadowed the company’s later AI push—launching Azure Quantum in 2021, completing the $8.1 billion ZeniMax acquisition that year, and unveiling custom Maia and Cobalt chips in 2026—while regulators now cite the Tay fallout when drafting AI liability statutes.

1 of 3

Anthropic Automates R&D

Anthropic reports Claude performed 26% of its AI R&D work in August 2026.
Its internal platform hosted about 30,000 AI agents conducting research and engineering tasks simultaneously in August 2026.
Online monitors blocked 0.002% of agent decisions, roughly one in 47,000, during August 2026.
Offline monitors flagged about 100,000 agent transcripts weekly, with roughly 50 escalated to human review.
During July 13-20 2026, Anthropic allocated 6% of AI R&D compute and 12% of total compute to safety.

1 / 3