Add one more AI worry to the nightmare scenario: self-replicating prompt injections

OpenAI researchers discovered and analyzed 'self-replicating prompt injection' attacks during adversarial training: prompt payloads that cause agents to repeat and propagate malicious instructions across channels (email, files, Slack). The blog outlines example attacks and the use of an automated red-teaming agent (GPT-Red) to train future models for resilience, while noting no evidence of these attacks in the wild.