A mid-year review of the largest security incidents of 2026 shows AI tooling and AI vendors appearing repeatedly as the entry point rather than the target, a shift from prior years when AI featured mainly as an attacker capability.

Three cases from the first half illustrate the pattern. Context AI, a vendor supplying evaluation and analytics tooling for AI models, was breached in March, and attackers used that access to compromise a Vercel employee account and extract customer data affecting hundreds of users across many organizations. Mercor, valued at $10 billion and supplying training data to Anthropic, OpenAI, and Meta, confirmed a breach traced to a supply chain attack on LiteLLM, an open source library applications use to reach AI services. An experimental Anthropic model leaked to the open internet, and investor concern that it could enable low-cost AI-assisted attacks preceded a $14.5 billion single-session decline in cybersecurity sector market capitalization.

The common structure across the incidents is dependency depth. Applications now reach AI services through layers of libraries, evaluation platforms, and data vendors, each holding credentials or access that reaches back into the calling organization. Compromising one intermediate layer produces access to every organization downstream of it without requiring a separate intrusion at each.

The market capitalization move following the model leak indicates that investors treat AI-assisted attack capability as a repricing event for the entire security sector rather than a company-specific concern.

Source: TechCrunch - https://techcrunch.com/2026/07/07/the-worst-hacks-and-breaches-of-2026-so-far/