Observed arrival · 2026-10-14
Ulterior Safety looks inside AI models for harmful intent
A solo-founder project says it monitors models’ internal activity to flag harmful requests before they reach the output, including when safety training has been removed.
- For
- AI researchers, funders, and teams deploying models
- Worth noticing
- The founder says he recorded internal activity across 19,000 replies; the displayed benchmark results are labeled small and still expanding.
Field notes
The page lays out two evaluation measures—harmful requests blocked and safe questions wrongly refused—and compares the project’s detector with the model’s original safety setup. Its reported tests use Qwen3.5-9B on JailbreakBench and XSTest, which the project calls small benchmarks that are still being expanded. The founder says he built the pipeline and recorded internal activity across 19,000 replies; the page presents no independent validation.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue