Cached tokens are prompt tokens that an AI platform stores and reuses when the same or similar text appears again. Instead of reprocessing the full prompt every time, the system can serve repeated material from cache, which may reduce latency and lower usage charges or track those tokens separately for billing.
In cyber security, caching is useful but sensitive. It can expose which prompts are being repeated, preserve data longer than intended, and create cross-tenant or stale-context risks if isolation is weak. Attackers may try to abuse prompt caching to infer prior inputs, replay expensive requests, or poison shared caches with misleading context. Defenders should isolate caches by tenant, expire entries quickly, avoid caching secrets or regulated data, and monitor unusual reuse patterns and billing spikes.