Workers AI fixes GLM-5.3 cache exposure
Cloudflare disclosed that 155 GLM-5.3 responses on Workers AI contained data from other customers after a context-length mismatch triggered an SGLang disaggregated-serving bug. Cloudflare took the model offline, patched its deployment, and upstreamed fixes; it found no evidence of deliberate exploitation.
This was an inference-infrastructure isolation failure, not a model flaw: a known KV-capacity mismatch met an engine bug that failed to stop rejected requests before handoff.
- –Prompts above 681,280 tokens could access stale KV pages from earlier sessions
- –Cloudflare’s testing checked status codes and performance, but not output isolation or coherence
- –Operators should compare actual KV-pool capacity with configured context length before every deployment change
- –SGLang users should upgrade to v0.5.22 or backport the admission, lifecycle, transfer, and corruption-detection fixes
- –Page ownership checks and formal verification may provide stronger protection than configuration validation alone
DISCOVERED
1h ago
2026-10-26
PUBLISHED
1h ago
2026-10-26
RELEVANCE
AUTHOR
Cloudflare