Cloudflare has disclosed a tenant isolation vulnerability affecting the GLM-5.3 Workers AI model, which accidentally exposed fragments of customer data across requests between late August and mid-September 2026. The issue, which was swiftly remediated following a security report, stemmed from a combination of a local resource configuration mismatch and a downstream bug within SGLang, the open-source inference engine utilized in Cloudflare’s disaggregated architecture. According to the company’s technical disclosures, no other model hosted on Workers AI was impacted, and an extensive post-incident audit revealed no evidence of deliberate exploitation by malicious actors.
The security issue came to light on September 8, 2026, when security researcher Gadi Evron submitted an alert through Cloudflare’s HackerOne program. Evron noted that extremely long prompts sent to the GLM-5.3 model on Workers AI occasionally returned text belonging to other users’ active sessions. Cloudflare engineers verified the report the following morning, declared a formal security incident, and successfully reproduced the data leak using two independent accounts within a 25-minute window. To contain the vulnerability, the affected model was taken offline less than 40 minutes after the incident declaration, redirecting incoming requests to return error responses while engineers investigated the root cause.
Between August 27 and September 9, 155 responses were affected globally. However, every single impacted interaction involved prompts exceeding 681,280 tokens. An analysis of the request timings and volumes confirmed that the occurrences matched ordinary long-context usage patterns rather than repetitive, targeted exploitation attempts. Service was restored later that same day after deploying a corrected configuration and a patched inference engine. Cloudflare published its comprehensive post-incident analysis concurrently with a technical disclosure from SGLang, detailing the timeline, the precise nature of the underlying software bugs, and the aggressive hardening protocols subsequently rolled out across all production environments.
The root cause of the cross-request key-value (KV) cache exposure required two distinct conditions to overlap. First, a resource misconfiguration occurred during the initial deployment of the text generation model developed by Z.ai. Cloudflare configured GLM-5.3 to utilize a 1,048,576-token context window, but the physical KV cache pool allocated on the prefill GPUs was only provisioned with enough physical memory to accommodate 681,280 tokens. Engineering teams were aware of this sizing discrepancy prior to launch, operating under the assumption that the inference engine would gracefully reject or return an error for any prompt exceeding the capacity limit. Because standard pre-launch validation tests consistently returned successful responses for long-context queries, the team did not recognize the configuration gap as an immediate risk.
The second and more critical factor involved a software bug in SGLang’s handling of over-length requests when operating in disaggregated prefill and decode mode. In this operational architecture, a dedicated pool of prefill GPUs processes incoming prompts to build the model’s KV cache, subsequently handing off the cache via remote direct memory access using Mooncake technology to a separate pool of decode GPUs responsible for generating the final text response. When a prompt arrived that exceeded the physical capacity of the prefill KV cache pool, SGLang’s internal scheduler correctly rejected the request. However, instead of cleanly terminating the process and returning a client error, the scheduler replaced the prompt with a one-token placeholder to allow it to pass through normal batch processing without triggering memory allocation errors on single-node setups.

In Cloudflare’s disaggregated environment, this mechanism broke down. The admission queues on both prefill and decode nodes monitored whether a request was marked as finished rather than checking if it had been rejected, meaning the aborted request continued to register as active. The prefill node generated a KV cache for the trivial one-token placeholder and transferred it to the decode node. Receiving a "transfer complete" signal, the decode node assumed the prefill process had overwritten the cache, when in reality it left the majority of the previous state untouched. Consequently, the decode engine generated responses by referencing stale cache pages that still contained residual data from earlier user sessions, resulting in data leakage.
To address the vulnerability before bringing the model back online, Cloudflare implemented immediate physical and software-level corrections. Engineers adjusted the GLM-5.3 resource configuration to match previous stable deployments, expanding the prefill KV pool capacity to roughly 2.04 million tokens—more than double the maximum advertised context window. Because valid prompts now effortlessly fit within the cache pool limits, the problematic rejection path is bypassed entirely during normal operations. Additionally, Cloudflare collaborated with SGLang maintainers to patch the engine, ensuring that any request carrying a pending rejection is halted at the disaggregation admission queue before any cache allocation or remote transfer can take place.
The incident has prompted broader engineering efforts to harden the infrastructure supporting large language models. The KV cache in GPU memory relies on shared pools of fixed-size pages that are rapidly overwritten between requests rather than zeroed out, a performance optimization common across modern AI inference engines. Because page ownership is dynamically tracked across disaggregated nodes, memory-safety bugs can introduce race conditions if a page is prematurely freed or read while an asynchronous transfer is still in flight. Cloudflare has since updated its internal deployment validation pipelines to automatically verify KV pool sizes directly from running engines before releasing any updates, moving away from relying solely on standard HTTP status responses.
Working closely with core engineers and maintainers at RadixArk, Cloudflare has contributed numerous upstream patches to the SGLang project. These updates include mechanisms to retire aborted prefill requests proactively, enforce per-transfer checksums so receiving nodes can validate KV cache integrity before generation begins, and ensure that request cancellations properly propagate across distributed system components. Furthermore, engineers are pursuing long-term architectural improvements, including prototype page-level tagging systems designed to detect and block read operations on unowned memory at the hardware and software interface, alongside formal verification methods to mathematically prove the safety of distributed RDMA transfer protocols.
Cloudflare has emphasized that transparency and collaboration within the open-source and security research communities remain essential for mitigating emerging threats in high-performance AI infrastructure. The company expressed gratitude to Gadi Evron for his clear, professional disclosure through HackerOne and commended the SGLang maintainers for their rapid response in deploying upstream fixes. Operators utilizing disaggregated SGLang deployments have been strongly encouraged to update their software dependencies to recent versions to incorporate these critical safety patches and ensure robust tenant isolation across all enterprise AI workloads.