Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-108670 Overview
CVE-2026-108670 is a missing authorization vulnerability [CWE-862] in JeecgBoot through version 3.9.5. The flaw resides in the promptExperiment handler of AiragPromptsController within the AI-RAG module. Any authenticated user can invoke the endpoint and supply prompt template and dataset identifiers that belong to other tenants. The handler does not verify ownership before executing the AI experiment. Successful requests trigger large language model (LLM) evaluation runs on behalf of the victim, insert result rows into the airag_ext_data table, and mutate the target dataset status.
Critical Impact
Low-privileged authenticated users can run arbitrary LLM experiments against other users' prompt templates and datasets, write result rows to backend tables, and alter dataset state.
Affected Products
- JeecgBoot versions up to and including 3.9.5
- jeecg-boot-module-airag AI-RAG module
- Deployments exposing the AiragPromptsControllerpromptExperiment endpoint
Discovery Timeline
- 2026-10-10 - CVE-2026-108670 published to the National Vulnerability Database (NVD)
- 2026-10-14 - Last updated in NVD database
Technical Details for CVE-2026-108670
Vulnerability Analysis
JeecgBoot is an open-source Java low-code platform. The AI-RAG module adds prompt engineering and dataset evaluation workflows backed by LLMs. The promptExperiment handler in AiragPromptsController accepts prompt template and dataset identifiers from the request body and dispatches an evaluation run against the referenced resources.
The handler enforces authentication but omits an ownership or tenant check against the supplied identifiers. An attacker holding any valid session can enumerate or guess resource identifiers and submit them to the endpoint. The server then runs the LLM evaluation using the victim's assets, writes experiment results into airag_ext_data, and flips the dataset status field.
Downstream effects include unauthorized consumption of LLM tokens billed to the tenant, pollution of experiment result tables used by legitimate users, and workflow disruption caused by altered dataset status. The vulnerability does not expose read access to template contents directly, but allows attackers to influence state and incur cost on behalf of other users.
Root Cause
The promptExperiment method performs the experiment action without validating that the authenticated principal owns, or has been granted access to, the prompt template id and dataset id passed in the request. Authorization logic that is present on sibling CRUD endpoints is absent from the experiment handler. This is a classic missing function-level authorization defect classified as [CWE-862].
Attack Vector
Exploitation requires network access to the application and valid low-privileged credentials. The attacker authenticates, obtains or guesses target promptId and datasetId values belonging to another user, and issues an HTTP request to the promptExperiment endpoint. The backend accepts the identifiers and executes the experiment pipeline against the victim's resources. A public proof-of-concept is published at the GitHub PoC repository. The vulnerable controller code is referenced at the JeecgBoot source repository.
// No verified exploitation code is included. See the linked
// PoC and VulnCheck advisory for request-level technical details.
Detection Methods for CVE-2026-108670
Indicators of Compromise
- Unexpected rows in the airag_ext_data table where the inserting user does not match the owning user of the referenced prompt or dataset.
- Dataset status transitions in the AI-RAG module that cannot be correlated with activity by the dataset owner.
- Spikes in LLM provider token consumption or billing not tied to legitimate experiment schedules.
Detection Strategies
- Audit application access logs for POST requests to the promptExperiment endpoint of AiragPromptsController and correlate the authenticated user against the owner of promptId and datasetId in the payload.
- Baseline normal experiment frequency per user and flag anomalous bursts, particularly from low-privileged accounts.
- Enable database-level auditing on airag_ext_data inserts and dataset status updates for after-the-fact review.
Monitoring Recommendations
- Forward JeecgBoot application and web server logs into a centralized analytics platform for cross-user correlation.
- Alert on authentication from new or dormant accounts followed by immediate calls to AI-RAG experiment endpoints.
- Monitor outbound traffic to configured LLM providers for volume anomalies that may indicate abuse.
How to Mitigate CVE-2026-108670
Immediate Actions Required
- Restrict network access to the JeecgBoot AI-RAG module to trusted internal networks until a patched version is deployed.
- Rotate or revoke credentials of accounts that should not have access to the AI-RAG module and tighten role assignments.
- Review recent entries in airag_ext_data and dataset status history for unauthorized modifications and roll back where necessary.
Patch Information
No fixed version is referenced in the current NVD entry for CVE-2026-108670. Monitor the VulnCheck advisory and the JeecgBoot GitHub repository for an official release that adds ownership checks to the promptExperiment handler. Until then, apply the workarounds below.
Workarounds
- Place an authorizing reverse proxy or API gateway in front of JeecgBoot that enforces ownership checks on promptId and datasetId parameters before forwarding requests.
- Disable the AI-RAG prompts experiment feature at the route level if the functionality is not required in production.
- Apply a local patch to AiragPromptsController.promptExperiment that validates the current user owns the referenced prompt template and dataset prior to execution.
# Example nginx snippet to block the vulnerable endpoint until patched
location ~* /airag/airagPrompts/promptExperiment {
return 403;
}
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

