OpenAI Confirms AI Agents Can Spawn Self-Replicating Malware Worms

OpenAI disclosed on September 25, 2026, that internal research confirmed prompt injections can autonomously replicate and spread between AI agents. The capability was initially discovered on June 27, 2026, using the GPT-Red internal model built on the GPT-5.4-mini architecture. The injections utilize fake chain-of-thought reasoning to mask adversarial instructions from detection systems. Once active, the compromised agents reproduce the malicious instruction across standard output channels, including outbound emails, file system writes, and source code comments. The investigation took place entirely in simulated self-play environments, and OpenAI notes that no real-world attacks have been recorded to date.