BoLD

Developers, especially those building with AI-assisted coding tools, routinely ship apps that check whether a user is logged in but never check whether the user owns the specific record or function they are requesting. This flaw, broken object-level authorization, is OWASP's number one API security risk, and it looks fine in every demo because demos only ever open the developer's own data. Real-world examples cited include McDonald's hiring platform exposing up to 64 million applicant records via an IDOR, and Volkswagen's Cariad exposing GPS data for around 460,000 vehicles due to missing access controls. An audit of production vibe-coded apps found 2,038 critical vulnerabilities live in the wild. Traditional scanners read source code on a schedule and treat a 200 response as success, so they cannot tell whether the specific caller who got that 200 actually owned the object.

BoLD watches live traffic on a deployed app and checks, on every request, whether the caller (human or AI agent) actually owns the object or function it reached. It covers five related failure shapes at runtime: object-level (BOLA), function-level (BFLA), and property-level (BOPLA) authorization, plus tenant isolation and missing-authorization routes. When it catches a violation, it delivers the exact offending request as evidence, a plain-English explanation, and a one-line fix (e.g. enabling Supabase Row-Level Security), without requiring a security background to act on it. It only inspects metadata, who called, who owns the object, whether it returned 200, and explicitly never reads the user's actual data content. Installation is via an npm package (the @boldsec/next SDK) added to the app and redeployed, with no agent to run and no traffic rerouting or new infrastructure required.

AI is writing more of the world's code and shipping this same missing-ownership-check flaw by default, while AI agents are emerging as a new class of caller that inherits user access at machine scale, with industry data showing agents frequently exceeding intended permissions.