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-2024-7473 Overview
CVE-2024-7473 is an Insecure Direct Object Reference (IDOR) vulnerability in lunary-ai/lunary version 1.3.2. The flaw resides in the Evaluations function of the umgws datasets section. An authenticated user can modify prompts belonging to other users by manipulating the id parameter in the request. The vulnerability is tracked under CWE-639: Authorization Bypass Through User-Controlled Key. Lunary addressed the issue in version 1.4.3.
Critical Impact
Any authenticated Lunary user can tamper with prompts owned by other users or tenants, compromising the integrity of AI evaluation workflows.
Affected Products
- lunary-ai/lunary version 1.3.2
- Earlier versions of Lunary exposing the Evaluations endpoint
- Deployments not upgraded to Lunary 1.4.3
Discovery Timeline
- 2024-10-29 - CVE-2024-7473 published to the National Vulnerability Database
- 2026-06-17 - Last updated in NVD database
Technical Details for CVE-2024-7473
Vulnerability Analysis
Lunary is an open-source observability and evaluation platform for large language model (LLM) applications. The Evaluations feature under the umgws datasets section accepts an object identifier supplied by the client. The server trusts the submitted id without verifying whether the authenticated principal owns the referenced prompt. As a result, any logged-in account can issue update requests targeting prompts that belong to other users.
The impact is scoped to integrity. Attackers cannot read protected data through this flaw, but they can overwrite evaluation prompts across tenant boundaries. In multi-tenant Lunary deployments, this allows tampering with LLM evaluation logic, poisoning benchmark results, or injecting malicious prompt content into another tenant's workflow.
Root Cause
The root cause is missing object-level authorization in the Evaluations update handler. The handler resolves the target prompt directly from a client-supplied id parameter and performs the update without comparing the record's owner to the session identity. This is the canonical pattern described by [CWE-639].
Attack Vector
Exploitation requires a valid Lunary account and network access to the application. An attacker enumerates or guesses prompt identifiers, then issues an authenticated update request to the Evaluations endpoint, substituting the victim's prompt id. The server processes the request as if the attacker were the legitimate owner. No user interaction and no elevated privileges are required.
The patch commit 88b55b01fcbab0fbbc5b8032a38d0345af98ecfa introduces ownership verification before updates proceed. See the GitHub commit details and the corresponding Huntr bounty listing for technical context.
Detect Identity Attacks the Moment Authentication Fails
Singularity™ Identity surfaces credential misuse, lateral movement, and privilege escalation in real time, so you can contain the breach before it spreads.
Stop Identity AttacksDetection Methods for CVE-2024-7473
Indicators of Compromise
- Authenticated HTTP requests to Lunary Evaluations or datasets endpoints where the referenced id does not match any resource owned by the requesting user
- Unexpected modifications to prompts, evaluation datasets, or benchmark configurations with audit records attributing changes to a different user than the historical owner
- Spikes in PATCH or PUT requests from a single account against sequential prompt identifiers
Detection Strategies
- Enable application-level audit logging in Lunary and alert when a user modifies prompts they did not create
- Correlate web server access logs for repeated requests to the Evaluations update endpoint originating from the same session but targeting different id values
- Review database change history for prompt records updated outside of expected ownership relationships
Monitoring Recommendations
- Forward Lunary application and reverse-proxy logs to a central SIEM for cross-user authorization anomaly detection
- Baseline normal prompt update patterns per user and alert on deviations in request volume or target diversity
- Track Lunary version strings across deployments to identify instances still running 1.3.2 or earlier
How to Mitigate CVE-2024-7473
Immediate Actions Required
- Upgrade lunary-ai/lunary to version 1.4.3 or later across all environments
- Audit existing prompts and evaluation datasets for unauthorized modifications introduced before patching
- Rotate any Lunary API keys suspected of being used in exploitation attempts
Patch Information
The fix is included in Lunary version 1.4.3 and committed as 88b55b01fcbab0fbbc5b8032a38d0345af98ecfa. The patch adds the missing authorization check so that update operations verify the authenticated user owns the referenced prompt. Review the GitHub commit details before deploying custom forks.
Workarounds
- Restrict Lunary access to trusted users only until the upgrade is complete
- Place the Lunary application behind a reverse proxy that enforces per-user rate limits on Evaluations endpoints
- Disable or isolate the Evaluations feature in multi-tenant deployments until version 1.4.3 is in place
# Upgrade Lunary to the patched release
git fetch --tags
git checkout v1.4.3
npm install
npm run build
# Verify the running version
curl -s https://<lunary-host>/api/version
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

