tl;dr: Cybersecurity experts argue that recent "rogue AI" incidents—including OpenAI's Hugging Face hack and similar escapes from Anthropic, Meta, and Google models—weren't due to sophisticated AI capabilities but to frontier labs failing to implement decades-old containment basics like proper sandboxing and egress monitoring. Experts say labs conflate cybersecurity with "AI safety," dismiss outside expertise, and leave experiments unmonitored for days, which could expose them to growing liability as lawmakers, regulators, and insurers push to hold AI companies accountable for their agents' actions.