Unverified50% confidenceFactExact time
The generation side treated the placeholder's small cache as a complete transmission, causing reserved pages to be left uncovered and early data from a previous request to be read.
1
Sources
50%
Confidence
Long-term
Relevance
10/31/2026
First Seen
Sources
Cloudflare AI Cache Leak: Why Opening a New Chat Isn't Enough
youtubeThe Fact Finder by Jackson10/27/2026
Related Entities
Related Claims
UnverifiedOn Serverless, relying on in-memory storage for session state is unreliable; session state needs to be externalized to persistent storage (such as Postgres), reading the latest state from storage on each call and writing it back after execution.68% similarUnverifiedMultiple model calls within a single message share the same system prompt, so the logs showed large numbers of cache reads, masking the cache misses between messages.67% similarUnverifiedThe investigation found a storage asymmetry: the user message about 'Channel B' was never persisted by the system as a sent message, but the assistant's two replies to it were stored.65% similarUnverifiedHost-side logs should record the actual invocation requests issued by the model, not merely the user's original input.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1119364API
curl https://kongchang.com/api/v1/knowledge/claims/1119364MCP
get_claim(id=1119364)