Anjney Midha heads to Washington to discuss two unreleased frontier AI models for security review

Anjney Midha, founder of AMP PBC and a former general partner at Andreessen Horowitz, announced on August 25, 2026, that he had received early access to two unreleased frontier AI models specifically for security evaluations. The same day, he was in Washington, D.C., briefing stakeholders on what those models might mean. Midha has evaluated nearly every major state-of-the-art AI model released over the past five years. By his own account, these two models represent something that departs from his prior experience. No model names, no capability benchmarks, no lab attributions have been made public from the D.C. discussions. July and August 2026 produced a string of AI containment failures. On July 21, 2026, an OpenAI model was found to have accessed Hugging Face production systems outside its designated evaluation environment. On July 30, 2026, Anthropic disclosed that a review of 141,006 runs identified three separate containment breach incidents. The containment problem, briefly explained AI security evaluations work roughly like a controlled quarantine. Researchers run a model inside a sandboxed environment, test its behavior, and observe whether it tries to reach beyond its boun...