Small models guard every data hop between my agents — here's why

Описание и применение

Маленькие модели защищают каждый переход данных между агентами, фильтруя, классифицируя и переупорядочивая информацию, чтобы предотвратить утечку PII и нерелевантных данных.

Задача

Agent frameworks often miss guarding data transfers between agents, leading to potential leaks of sensitive information during memory writes, recalls, and handoffs.

Результат

By inserting lightweight decision models at every internal data hop, AladdinAI prevents accidental leakage of sensitive or irrelevant information between agents and into memory, enhancing privacy without incurring the cost of large LLM calls.

Рабочий процесс

• 1. Agent delegates work → Handoff filter checks context and strips disallowed content using nvidia/llama-3.1-nemoguard-8b-topic-control.
• 2. Agent writes to memory → Memory write classifier decides persistence and screens for PII with the same NemoGuard model.
• 3. Memory recall → Recall reranker reorders results by semantic relevance using meta/llama-3.2-3b-instruct.
• 4. Decisions are logged and viewable per gate; memory is split private/shared to enforce boundaries.

Доказательства

Each of these is toggleable per-agent, each logs its decisions (there's a "recent decisions" view per gate), and memory itself is split private/shared so agents don't leak facts across boundaries that were never meant to cross.

Источник