The paper shows how to build tiny, fast safety checkers (called probes) that look inside a big AI’s brain activity to spot dangerous cyber-attack requests.
Memory-T1 teaches chatty AI agents to keep track of when things happened across many conversations.