CVE-2026-59704: Cap Video AI Metadata Authorization Bypass – Sensitive Data & Credit Theft Risk

Technical summary

CVE-2026-59704 is an authorization bypass vulnerability in Cap's GET /api/video/ai endpoint. The endpoint performs authentication checks (requires a valid session) but omits ownership and membership validation before serving private video AI metadata and accepting AI generation requests. An attacker supplies a target video ID in the request; the server returns AI-generated summaries, chapters, and titles without confirming the requester owns or has access rights to that video. Additionally, the endpoint processes AI generation jobs and charges the video owner's credit account without verifying the request originated from an authorized party. The vulnerability is driven by missing access control logic (CWE-862).

Business impact

This vulnerability creates two distinct business harms. First, proprietary or confidential AI-generated video metadata (summaries, chapter breakdowns, titles) can be exfiltrated by any authenticated Cap user, potentially exposing competitive intelligence or sensitive content analysis. Second, attackers can programmatically drain video owner credit balances by triggering repeated AI generation requests, inflating customer support costs and undermining trust in the platform's billing integrity. Organizations using Cap for transcription or video summarization of sensitive material face data leakage and unexpected operational costs.

Affected systems

Cap's video AI metadata endpoint (GET /api/video/ai) is the affected component. Any instance of Cap where users have authentication credentials and the AI metadata feature is enabled is vulnerable. The issue affects all users whose video content can be queried by other authenticated users, regardless of sharing or access policy settings.

Exploitability

Exploitability is straightforward. The vulnerability requires only valid Cap authentication credentials (PR:L in the CVSS vector), no user interaction, and network access to the API. Attack complexity is low: an attacker simply needs to enumerate video IDs—either by iterating sequentially, scraping public references, or leveraging information disclosure elsewhere—and request metadata for each. Automated tooling can scale this to harvest metadata or exhaust credits across hundreds or thousands of videos in minutes. No special network position or platform-level access is required.

Remediation

Cap must implement authorization checks in the GET /api/video/ai endpoint before returning metadata or accepting AI generation requests. For each request, verify that the authenticated user either owns the video, has explicit sharing permissions, or belongs to a team/workspace with collective access rights. AI generation requests must additionally validate that the requester is authorized to consume credits on behalf of the video owner. Deploy these checks server-side; do not rely on client-side filtering. Review and test all other API endpoints that expose video metadata, summaries, or user content to ensure similar gaps do not exist elsewhere.

Patch guidance

Check Cap's advisory channels and release notes for a patched version addressing authorization validation on the /api/video/ai endpoint. Prioritize deployment of the patch to all Cap instances, particularly those serving sensitive or confidential video content. Verify the patch includes both metadata access controls and credit consumption authorization before considering the vulnerability resolved. If you are a Cap customer, contact your vendor immediately for patch availability and timeline.

Detection guidance

Monitor API access logs for repeated requests to GET /api/video/ai with varying video IDs from a single user or IP address, especially if those video IDs are not owned by that user (cross-reference against your video ownership records). Alert on any unusual patterns of AI generation job submissions that correspond to videos a user does not own. Review CloudTrail or equivalent audit logs for unexpected API calls to this endpoint. Implement rate limiting on /api/video/ai requests per user or API key to slow brute-force enumeration of video IDs. Consider temporarily disabling or restricting the AI metadata endpoint until a patch is available and deployed.

Why prioritize this

HIGH severity: The vulnerability combines confidentiality (CWE-862 authorization bypass enables private metadata exfiltration) with financial impact (credit theft). Authenticated access requirement (PR:L) and low attack complexity make this a practical threat that requires neither special network position nor user interaction. Organizations storing sensitive video analysis or maintaining tight credit budgets should patch urgently. Not currently listed on CISA KEV, but should still be treated as a near-term priority.

Risk score, explained

CVSS 7.1 (HIGH) reflects high confidentiality impact (sensitive AI metadata disclosure), low integrity impact (credit deduction without authorization), and no availability impact. The attack vector is network-based, complexity is low, and only low privileges (valid authentication) are required. The scope is unchanged (the user gains unauthorized access within the same application). This score appropriately captures the threat of data exfiltration and financial abuse without overstating denial-of-service risk.

Frequently asked questions