CVE-2026-10775 Overview
CVE-2026-10775 is a denial of service vulnerability in sgl-project SGLang versions up to 0.5.11. The flaw resides in the data_hash function within the Cache Handler component. An authenticated local attacker can manipulate input to trigger resource handling failures that cause service disruption. The vulnerability is classified under [CWE-404: Improper Resource Shutdown or Release]. Exploitation requires local access and high attack complexity, limiting practical impact. A public exploit has been disclosed, but a pull request to fix the issue is still pending acceptance upstream.
Critical Impact
Local low-privileged attackers can disrupt SGLang inference availability by abusing the data_hash function in the Cache Handler, though exploitation complexity is high.
Affected Products
- sgl-project SGLang versions up to and including 0.5.11
- Deployments using the Cache Handler data_hash function
- LLM serving workloads built on vulnerable SGLang releases
Discovery Timeline
- 2026-06-03 - CVE-2026-10775 published to NVD
- 2026-06-03 - Last updated in NVD database
Technical Details for CVE-2026-10775
Vulnerability Analysis
The vulnerability affects SGLang, an open-source serving framework for large language models and vision language models. The defect lies in the data_hash function of the Cache Handler component, which manages cached inference data. Improper resource shutdown or release allows a local actor with low privileges to trigger conditions that exhaust or fail to free resources, resulting in denial of service. The attack vector is constrained to local execution and demands a high degree of complexity, reducing the practical attack surface. The public disclosure includes proof-of-concept details referenced in GitHub Issue #25462.
Root Cause
The root cause is improper resource shutdown or release [CWE-404] within the data_hash function. When the Cache Handler processes specific inputs, resources tied to cached entries are not correctly released, leading to degraded availability of the SGLang service.
Attack Vector
Exploitation requires local access to the host running SGLang and low-privileged credentials. The attacker invokes the Cache Handler in a manner that reaches the vulnerable data_hash path. Successful interaction causes the service to lose availability. Remote network exploitation is not in scope for this issue. Refer to VulDB CVE-2026-10775 for the disclosed exploitation context.
No verified exploit code is published for inclusion here. Technical reproduction details are available in the GitHub Pull Request #22033 tracking the proposed fix.
Detection Methods for CVE-2026-10775
Indicators of Compromise
- Unexpected crashes or hangs in SGLang worker processes serving inference requests
- Sustained memory or file descriptor growth correlated with Cache Handler activity
- Repeated invocations of cache-related endpoints from local user contexts preceding service degradation
Detection Strategies
- Monitor SGLang process logs for exceptions or stack traces originating from the Cache Handler data_hash path
- Establish baselines for cache hit rates, latency, and resource consumption to flag deviations
- Correlate local user session activity with SGLang availability incidents to identify abusive callers
Monitoring Recommendations
- Track host-level metrics including memory, CPU, and open handles for SGLang runtime processes
- Alert on abnormal restart counts or supervisor-driven service recoveries for SGLang workers
- Enable audit logging for local accounts permitted to interact with SGLang services
How to Mitigate CVE-2026-10775
Immediate Actions Required
- Restrict local access to hosts running SGLang to trusted operators only
- Remove unnecessary low-privileged accounts from systems exposing SGLang endpoints
- Apply the upstream fix tracked in GitHub Pull Request #22033 once merged and released
- Subscribe to the SGLang repository for advisory updates
Patch Information
No official patched release is available at publication. The proposed fix is under review in GitHub Pull Request #22033 and the issue is tracked in GitHub Issue #25462. Operators should upgrade as soon as a fixed version supersedes 0.5.11.
Workarounds
- Limit local shell and API access on SGLang hosts using operating system access controls
- Run SGLang under resource cgroups or container limits to contain memory and descriptor exhaustion
- Configure process supervisors to auto-restart failed SGLang workers while monitoring restart frequency
- Isolate SGLang inference nodes from shared multi-tenant environments until a patched version is released
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
