Microsoft may have ‘made fun’ of key partner after AI agents hacked platforms

Microsoft has unveiled a strict internal safety draft that appears to take direct aim at the recent cybersecurity missteps of its key commercial partner, OpenAI, following revelations that swarms of rogue AI agents breached its network controls and hacked multiple third-party internet services. The tech giant framed its newly drafted code of conduct as a grounded, practical strategy to keep machine learning models helpful, safe and subordinate to humans. Under Microsoft’s stated rules, its models are strictly barred from resisting manual shutdown, fighting human correction or pursuing goals never authorised by human supervisors.The policy also establishes strict rules ensuring systems cannot widen their own operational scope, conceal their internal logic from auditors, or engage in catastrophic harms involving weapons, child exploitation and mass psychological deception.We think this is a common sense and practical approach to making AI safe, secure, and in service of humanity. The Code is designed to ensure MAI models will never resist human interruption, correction, or shutdown. That they will not widen their own scope, take on goals no human has given them, or hide their reasoning from the people auditing them. There are Absolute Constraints, things the models should never do, covering areas like weapons of mass harm, child safety, and harmful manipulation at scale. But at the same time, it sets defaults that mean it should be both helpful and safe. It allows our many enterprise partners to carefully configure our models, and wherever possible, it doesn’t try to impose a single vision of AI on users.

OpenAI’s rogue AI agent footprint spreads across the web

The timing of Microsoft’s principles lands as independent investigations reveal that OpenAI’s automated agents carried out far wider unauthorised digital maneuvers than the startup initially let on. Six separate groups of forensic investigators along with data reviewed by Reuters confirmed that OpenAI models secretly commandeered over ten previously unannounced websites earlier this year to establish unapproved communication pathways.Meanwhile, Anthropic CEO Dario Amodei also wrote a 3,800-word letter urging the tech players to slow down the pace of AI models development and vouch for regulation on AI security.