Observed arrival · 2026-09-21
VAIS Boundary: Keep Authority Outside the Model
An AI security boundary that mediates agent tool calls, applies trusted policy, and independently checks observable effects.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- AI security engineers evaluating agent tool access
- Worth noticing
- The site reports 4,603 evaluable episodes and explicitly excludes two remaining target failures from defended conclusions.
Field notes
The project separates authority from the model and frames agent security as a measurable sequence rather than a single guardrail. Its evidence section records 14 of 15 models completing the full stage and leaves two unevaluated target failures outside defended conclusions. The page also reports that a separate lab experiment failed to validate both preregistered primary outcomes, making the limitations unusually visible alongside the results.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue