AI Risks Spur Safety and Governance Concerns

AI’s shift from chatbots to autonomous agents is raising concerns that safety measures and governance may not keep pace. OpenAI says it found self-replicating prompt-injection vulnerabilities in training environments, with no reported real-world incidents, and is using adversarial training to improve model resistance. A test-design guide warns that AI-generated tests can pass despite faulty code, while urging developers to check whether plausible defects would still leave tests green. Proposed rules in California and the European Union would monitor data centers’ water and energy use; analysts also say AI could reduce emissions by cutting food waste, excess inventory and inefficient transport. A Grant Thornton review found that fewer than 1% of directors at major Hong Kong-listed firms disclose AI or cybersecurity expertise, despite many identifying those areas as key risks.