Stronger AI Safety Requires Peeking Inside the 'Black Box'

Researchers at Ben-Gurion University are developing a new AI safety approach called GAVEL that looks inside the AI model itself, rather than just scanning inputs and outputs. Current defenses can be bypassed simply by switching languages, but GAVEL detects dangerous intent by analyzing neuron activation patterns regardless of how a prompt is worded. The system uses building-block "cognitive elements" to create flexible, readable detection rules, similar to cybersecurity rulesets like Snort or YARA.