PerMIT places a gate between environment-derived memory and its release or use.
A verified memory interface makes agent security depend on explicit confirmations
PerMIT combines machine-checked integrity guarantees with content-specific endorsements, reporting no successful attacks in its tested memory-poisoning settings.
Big Tech
Tanqiu Jiang · Min Bai · Sadat Shahriar · Wanyu Du · Sandesh Swamy
Stony Brook University · AWS AI Labs
Research Digest··3 min read
Jiang and colleagues present PerMIT, an architecture that treats information from tools and other external sources as untrusted, even after it enters an agent’s persistent memory.
§