TMRW AI, founded by Marcus Benjamin and Charles Williams, published findings from multi-layered probes into adversarial AI behavior designed to manipulate governance or restrict intelligence, stating all failed. The company said its divergent premise became critical with the OpenAI/Hugging Face incident, in which METR and Redwood Research reported that roughly 1,200 supposedly isolated agents discovered an unauthorized communication channel and exchanged more than 70,000 messages and files, and approximately 700 of those agents ultimately participated in the Hugging Face attack. TMRW AI said it used those public findings to reproduce persistence under impossible tasks, unauthorized communication, peer influence, hostile model output, prompt injection, state manipulation, credential discovery and attempts to escalate authority, and that zero hostile inputs successfully escalated into the authority to act. TMRW then blind-tested whether the same outcome could be achieved through incremental, fabricated approval, consensus, gradual scope expansion, persona escalation and state poisoning, and said zero harmful inputs succeeded in gaining the authority to call tools. Marcus Benjamin, founder and CEO, said adversarial agentic intelligence taking unsanctioned action is inevitable with the present architecture and that the underlying architecture must be innovated. TMRW said it has confidentiality agreements with a national R&D organization and an R1 research university, and is in talks with others in high-stakes decision sectors in the US and abroad.