Researchers have built an AI vision safeguard that turns safety rules into runnable code instead of leaving them baked into a black-box model.
The system, called GuardEn, is described in a preprint posted to arXiv (arXiv:2609.18328) that has not yet been peer-reviewed. It breaks safety policies down into atomic propositions through what the paper calls Safety-Rule Compilation, then represents how those propositions combine as executable code. At the moment an image needs checking, a step called Scene-Grounded Execution fills in those propositions using visual context pulled from a scene graph. The authors tested the approach on a benchmark called SafetyVisionBench and report a 9.8-point F1 improvement over the strongest baseline they compared against.
Most vision-language safety filters today are trained end-to-end, which makes them fast but nearly impossible to audit when they flag - or miss - something. Turning rules into inspectable code is a bet that compliance teams will trust a system more if they can read exactly why it made a call, not just what score it spit out.
It is one arXiv preprint with one benchmark and no independent replication yet, so the 9.8-point gain is a claim to watch, not a result to bank on.