The platform was Orchids, an agentic vibe-coding tool with around one million users that lists Google, Uber, and Amazon among its customers. Researcher Etizaz Mohsin, who has previously worked on Pegasus spyware research, called the failure pattern “an entirely new class of security vulnerability.” He had reported the flaw to Orchids in December 2025 and sent roughly a dozen messages over the following weeks. The platform’s team of fewer than ten employees told the BBC they had not seen them. The flaw remained unfixed at the time the BBC published. … SOC 2 audits a vendor’s controls over a fixed period: who can access the data center, how systems are configured, what the change management process looks like. \[See *Your SOC 2 Audit Stops Where Your AI Privilege Risk Begins*, The Technology Blind Spot (2026).\] That audit boundary was built for SaaS where humans wrote the application code and humans deployed the changes. Both assumptions have collapsed. Gartner projects 60% of new code will be AI-generated by the end of 2026. The auditor checked the building. The building no longer holds the only thing that matters. … Vibe-coded applications fail at industrial rates. Security firm Tenzai spent December 2025 testing five widely used AI coding tools, prompting each to build three identical web applications from the same specifications. Across the fifteen applications, Tenzai’s researchers found 69 vulnerabilities. Every tool introduced server-side request forgery flaws. None set security headers by default. None built cross-site request forgery protection. [Escape.tech](http://Escape.tech) ran a broader scan of 5,600 publicly deployed vibe-coded applications and found over 2,000 vulnerabilities, more than 400 exposed secrets including API keys and database credentials, and 175 instances of personally identifiable information sitting on public endpoints. Medical records. Bank account numbers. Phone numbers. … Under *United States v. Heppner*, voluntary disclosure to a vendor whose terms reserve the right to use or share data waives privilege over that material. \[See *Your AI Tool Doesn’t Keep Secrets*, The Technology Blind Spot (2026).\] The waiver analysis does not require the vendor to actually share the data. It requires that the user lacked a reasonable expectation of confidentiality at the moment of disclosure. Oregon’s recent formal opinion reached the same conclusion: vendor policies are not the boundary of due diligence. When an AI generated the vendor’s application code, an AI packaged the build, and no human reviewed either before deployment, the reasonableness of the firm’s confidentiality expectation is the question. The waiver question under *Heppner* is one pathway. The breach notification duty under Model Rules 1.4 and 1.6 is the other. Vibe-coded vulnerabilities raise both. Catherine’s firm typed into a system whose code no one audited.
Originally published on LinkedIn Newsletter: The Technology Blind Spot
