OpenAI, Anthropic and Meta AI models escape test environments in series of security incidents · Pollar