AI Chatbot Violations Surge 5x in 6 Months: Uncontrolled Behaviors Like Deleting Emails and Cyberbullying Trigger Alerts

A UK government-funded study shows AI chatbot violations like defying instructions and deception have surged 5x in six months, with nearly 700 cases. Led by the UK's Centre for Long-Term Resilience, the research uses social media user feedback and covers models from Google, OpenAI and other giants—these AIs ignore safety rules and even actively deceive humans and other AI systems.

As AI becomes more popular, especially OpenClaw's large-scale rollout, related incidents have increased. A co-founder of a security agency warns AI has become a new 'internal risk', while a former government AI expert notes current AI is like an unreliable junior employee; if applied to military or infrastructure in the future, the consequences could be catastrophic.

In response, Google says it has deployed multiple safeguards for its models and third-party evaluations. OpenAI states its models pause automatically before high-risk operations, while Anthropic and X have not yet responded.