/

CVE-2026-90603

Severity:

HIGH RISK

(7.3)

Unrestricted Upload (CWE-434) in Open-Generative-AI

CVE-2026-90603

Deferred

Vulnerability type

Category:

Unrestricted File Upload

CWE:

CWE-284, CWE-434

Impact:

Unauthorized File Upload

Severity:

HIGH RISK

(7.3)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

Open-Generative-AI contains an unrestricted file upload vulnerability in the S3 Upload functionality exposed through /api/upload-binary. If a remote attacker can reach the endpoint and influence the x-proxy-target-url value, they may upload files without adequate authorization or type restrictions.

Affected context

The issue affects deployments of Anil-matcha Open-Generative-AI through versions 1.0.11 and 2.0.0 when the S3 Upload endpoint is exposed and does not sufficiently constrain proxy target handling or uploaded content. Risk depends on how uploaded objects are stored, whether they are publicly reachable or executable, and the permissions available to the application.

Exploit preconditions

  • Attack surface: Network-accessible /api/upload-binary endpoint
  • Required access: Ability to send requests to the upload functionality; authentication requirements may vary by deployment
  • Trigger point: Processing of an upload request containing a user-controlled x-proxy-target-url value
  • Why exploitation is possible: Insufficient authorization and upload restrictions may allow attacker-controlled content to be written to an unintended or insufficiently protected destination

Why this happens

The vulnerability reflects inadequate access control and file upload validation. A request-controlled proxy target is accepted in the S3 Upload flow without reliably restricting where content may be sent or enforcing strong controls over the uploaded object. This maps primarily to CWE-434: Unrestricted Upload of File with Dangerous Type, with CWE-284: Improper Access Control also relevant where authorization boundaries are not properly enforced.

Potential impact

  • An attacker may store unauthorized files in application-managed or cloud-backed storage
  • Uploaded content could be used to host malicious material or phishing resources
  • Public or executable storage configurations could increase the risk of code execution or client-side attacks
  • The application or associated storage account could experience data integrity, availability, or reputational impact

General mitigation guidance

Mitigation generally focuses on enforcing authorization and treating every upload destination and file as untrusted.

Common defensive patterns include:

  • Allowing uploads only to server-selected destinations rather than accepting arbitrary proxy targets
  • Enforcing strict allowlists for permitted storage locations, buckets, prefixes, and endpoints
  • Validating file type, size, name, and content independently of client-supplied metadata
  • Storing uploads outside executable or publicly served paths by default
  • Applying least-privilege permissions to application and storage identities
  • Using isolated storage with safe content-disposition and access policies
  • Limiting upload rates and recording security-relevant upload events

How teams detect issues like this

Detection approaches can include:

  • Static analysis for upload handlers that use client-controlled paths, URLs, bucket names, or object keys
  • Code review checks for missing authorization, content validation, destination allowlists, and executable storage separation
  • Web-application policy rules that alert on unusual x-proxy-target-url values or uploads to unapproved destinations
  • Monitoring object-storage events for unexpected prefixes, file types, public-access changes, or anomalous upload volumes
  • Automated security review tools that identify unrestricted upload and access-control patterns during CI analysis

Sources

SourceLink
github.comgithub.com/Anil-matcha/Open-Generative-AI/
github.comgithub.com/Anil-matcha/Open-Generative-AI/commit/f013270957f75e439eaf97eb2a93decb32a4543e
github.comgithub.com/Anil-matcha/Open-Generative-AI/issues/310
vuldb.comvuldb.com/cve/CVE-2026-90603
vuldb.comvuldb.com/submit/914005
vuldb.comvuldb.com/vuln/403185
vuldb.comvuldb.com/vuln/403185/cti

Frequently asked

questions

Yes, the affected upload endpoint may be reachable remotely. Exploitability depends on network exposure, authentication requirements, and the deployment's upload and storage controls.

Not necessarily. The required access depends on whether the endpoint enforces authentication and authorization before processing upload requests.

Do not accept arbitrary upload destinations from clients. Select destinations server-side, enforce authorization and allowlists, validate content, and keep uploaded objects non-executable and private by default.

Combine secure upload design with least-privilege storage permissions, schema and content validation, endpoint access policies, object-storage monitoring, and review rules for client-controlled paths or URLs.

Find out what's already
exploitable in your codebase.