AI inference security: leaks, scrapers, and protocol sprawl
A post-mortem found cross-request KV cache exposure in the GLM-5.3 model on Workers AI, affecting 155 responses. One user said it might explain the unrelated responses they'd seen from a certain provider's API. Commenters liked the post-mortem for its ownership and its clear explanation of why testing missed the bug, and praised how fast the model was pulled. Others suspect other providers either don't know or don't care that this happens.
The pattern: shared inference infrastructure has failure modes that look like classic multi-tenant bugs, but the data leaking is prompts and answers. Meanwhile the plumbing is multiplying. Microsoft's Type Server Protocol joins LSP, DAP and MCP in the 'protocol of the week' list, and commenters complained that it should be called the Python Type Server Protocol.
The reaction was skepticism, not excitement. People are tired of new protocols and want fewer, better ones.
So what?
If you serve other companies' prompts, isolation between requests is a core product promise, so test it. If you buy inference, ask providers about cache isolation and incident disclosure. Choose boring, widely adopted protocols over novel ones.