PuzzleMask slips past quick AI gates 23 times

Check Point Research says its PuzzleMask plain-prose wrappers passed a quick LLM policy check in all 23 trials, while a stronger target recovered and acted on embedded payloads in 17 of 18 tests; the controlled study reports a lab result, not a production breach. The finding shows why a gatekeeper that only labels input cannot be a sole control when a target model has more reasoning or tool access. Teams building agents should add output and action monitoring plus durable sandbox and policy controls; OpenAI’s Astra access gate makes the same capability-tier point.