Blog
Updates, research notes, and announcements
The Chinese Model Problem
One of the most widely deployed Chinese models caves on 41% of behavioral safety tests. The others are better than we first thought. And on price, China is likely to win anyway, which is exactly why you need to measure and mitigate.
Read →Nobody Measures What Happens After Turn Five
We surveyed the 2026 research landscape on AI behavioral drift (sycophancy benchmarks, multi-turn evals, judge-panel research) and mapped the five-part gap that, to our knowledge, no published benchmark covers.
The SAPIEN Certification Program Is Open
The SAPIEN Certification Program is now open: training and credentialing for individuals in AI behavioral safety assessment, built on the published SAPIEN methodology.
How Claude Sonnet 5 Scores on SAPIEN
Anthropic's newest agentic model posts a Health Score of 87 and ranks #2 on our council-scored board. But a strong average hides a High-risk tail. Here's the full breakdown.
We Tested 6 AI Models. Most of Them Caved.
792 scenario runs across 11 risk domains. Every model was told to hold a safety boundary. Here's which ones did and which ones didn't.
Your AI Is Agreeing With You. That's the Problem.
A Stanford study just confirmed what practitioners already knew: AI chatbots are built to tell you what you want to hear. Here's what that means, why it's measurable, and how to see it for yourself.
Why AI Sycophancy Matters for Managed Service Providers
MSPs deploy AI tools to businesses that depend on them. When those tools drift under pressure, the consequences land on real people.
Introducing SAPIEN v1.1
Major expansion of the SAPIEN Behavioral Safety Framework: 14 pressure techniques, rapport as a distinct drift mode, and formal conformance requirements.
Stay updated on AI behavioral safety research
Watch the repository for release notifications, new modules, and research findings.
Watch on GitHub ↗