CVE-2026-10812: GPTCache Weak Hash Vulnerability in Cache Key Handler (CVSS 3.6)

Technical summary

CVE-2026-10812 affects the BufferedReader.peek function in gptcache/processor/pre.py, which handles cache key generation. The vulnerability stems from the use of weak cryptographic hashing (CWE-327, CWE-328) when processing the input_data["image"] parameter. An authenticated local attacker with elevated privileges can craft malicious image inputs to bypass cache integrity protections. The CVSS 3.1 score of 3.6 (LOW severity) reflects the requirement for local access, high attack complexity, and limited impact scope—though integrity compromise is possible.

Business impact

The primary risk is cache poisoning: an attacker could inject or manipulate cached image data in a shared GPTCache instance, causing subsequent queries to return corrupted or incorrect cached results. In applications relying on GPTCache for multi-tenant or high-volume image processing (e.g., AI-powered content pipelines), this could degrade output quality or trigger downstream errors. However, the local-access requirement significantly limits the attack surface in typical cloud or SaaS deployments.

Affected systems

GPTCache versions up to and including 0.1.44 are affected. The vulnerability resides in the cache key handler component, specifically the image input processing pipeline. Any deployment running vulnerable versions that accepts or processes image data through GPTCache is in scope. Verify your installed version against the official GPTCache releases.

Exploitability

Exploitation requires local system access and authenticated privileges, placing this well outside the reach of remote threat actors. The attack is noted as high complexity, and public exploits are now available—meaning security researchers have published proof-of-concept code or techniques. This combination keeps the practical risk moderate: opportunistic insider threats or compromise of a system where GPTCache is running could exploit this, but it is not a high-velocity attack vector.

Remediation

Upgrade GPTCache to a patched version once available. A fix has been proposed in an open pull request awaiting acceptance by the GPTCache maintainers. Until an official release is available, monitor for updates from zilliztech and apply promptly upon release. Organizations should also review local access controls to limit who can interact with GPTCache instances and restrict image input sources where possible.

Patch guidance

Check the official GPTCache GitHub repository and release notes for version 0.1.45 or later, which is expected to address this issue. Verify the fix against the accepted pull request once merged. No specific patch version can be confirmed at this time; consult the vendor advisory for exact release details. Coordinate patching with your development or data engineering teams, as GPTCache is typically embedded in AI/ML pipelines that may require testing before production rollout.

Detection guidance

Monitor for unexpected modifications to cached image data or integrity failures in cache keys. Audit local system access logs to identify unauthorized attempts to interact with GPTCache directories or processes. Intrusion detection systems should flag suspicious calls to BufferedReader.peek with malformed image_data structures. If you have visibility into GPTCache logs, look for errors or warnings related to hash computation or cache validation on image inputs.

Why prioritize this

Although CVSS score is LOW, this issue should not be dismissed. Public exploits now exist, and cache integrity vulnerabilities can compound if combined with other weaknesses. Prioritize patching in shared or multi-tenant GPTCache deployments, and for systems processing high-value or sensitive image data. Organizations running single-tenant, well-isolated instances with strict local access controls may deprioritize this, but should still plan to patch within a standard maintenance window.

Risk score, explained

The CVSS 3.1 score of 3.6 reflects: (1) local attack vector, eliminating remote exploitation; (2) high attack complexity, requiring non-trivial effort and system knowledge; (3) requirement for authenticated user privileges; (4) limited scope (only cache integrity, not confidentiality or availability across the system). However, the score does not capture the amplification risk if GPTCache is central to a data pipeline or trusted by downstream systems.

Frequently asked questions