Disclosed weakness · no CVE recorded
AgentScope could reuse embedding results from the wrong API endpoint
AgentScope · AgentScope · embedding cache
First documented disclosure · Review checkpoint
Summary
A cached vector from one OpenAI-compatible endpoint could be returned in place of a vector from another endpoint.
Severity and exploitation evidence
No CVSS score recorded in this review.
Exploitation not established: The reviewed project publications document the weakness. They report no known malicious use.
Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.
How the weakness works
The cache identified requests by model name and parameters while omitting the provider endpoint.
Conditions required
Two endpoints share an embedding cache and receive matching input and request parameters under the same model name.
Potential impact
An application could index or query data with vectors from a different model, compromising retrieval integrity.
Affected versions and configurations
Pre-fix implementation identified by the project. The full release range is unspecified.
Fix and mitigation status
2.0.12 includes the correction.
Practical next steps
Upgrade to AgentScope 2.0.12 or a later fixed version.
Operating lesson
Keep endpoint-specific caches separate until upgraded.
Evidence and limits
The public issue and patch document the prior behavior and its correction. AgentScope lists the fix in its November 16 stable release.
Sources
- AgentScope original issue
Primary source ·
- AgentScope correction and regression checks
Primary source ·
- AgentScope 2.0.12 release notes
Primary source ·