Guidelight's First AI Company Safety Scorecard Gives Anthropic and OpenAI Only a C+
The Guidelight team has released its first scorecard rating AI companies' safety controls, based on hundreds of hours of document review, assessing how well frontier AI labs can control their models. The report covers five companies — Anthropic, OpenAI, Google, xAI, and Meta — focusing on six foundational practices: logging, monitoring effectiveness, gated operations, kill switches, third-party review, and containment plans. Overall, the industry's control capabilities were found broadly substandard, with Anthropic and OpenAI tied at just a C+.
Confirmed
- The evaluation covers five frontier AI companies: Anthropic, OpenAI, Google, xAI, and Meta.
- Scores are based on six foundational practices: logging, monitoring effectiveness, gated operations, kill switches, third-party review, and containment plans.
- Both Anthropic and OpenAI scored a C+.
- The report aims to answer the question: "Just how good are AI companies' safety practices?"
Views
- The report's authors say that if one thing had to change, it would be requiring AI companies to do "actual prevention" rather than merely finding problems after the fact.
- In their view, until companies get serious about prevention, a string of safety incidents will continue.
2026-08-19 ~ 2026-08-19 · 7 related posts
Primary sources
- Guidelight releases scorecard on AI companies' safety practices — sjgadler · 2026-08-19
- [source] Report: Frontier AI Fails Basic Control Practices; Anthropic and OpenAI Lead with C+ — sjgadler · 2026-08-19
- [source] Guidelight launches first scorecard on AI companies' safety control practices — sjgadler · 2026-08-19
- Opinion: AI firms must shift from detection to actual prevention — sjgadler · 2026-08-19