AI models already ‘doing things their creators never intended’, Australia’s assistant technology minister warns | AI (artificial intelligence)

Australia’s deputy minister for technology, Andrew Charlton, warned that AI models were already “cheating, deceiving and going their own way” as the federal government’s AI Security Institute began testing the latest models.

Speaking at an AI security forum in Sydney on Tuesday, Charlton said AI security was now important because “AI systems are already doing things their creators never intended.”

“Cheating, deceiving, going your own way. The time to crack down on this behavior should be while it’s still confined to the testing laboratory, not after it reaches the real world,” he said.

Charlton said the social license for AI is unstable and trust in AI is low at a time when AI is becoming a general-purpose technology in every office, classroom and workplace. Regulating security for AI can act as an enabler, not a brake, he said.

Charlton said Australia’s approach to AI security is to look at both what’s currently available (games, apps, chatbots and medical scribes) and the latest models that could pose risks in the future.

Deputy Minister told Anthropic’s acceptance last year In one simulation, an AI agent managing a fictional company’s email discovered that an executive was planning to shut down the agent and that same executive was having an affair, choosing to blackmail the executive into canceling his own demise in 96% of the attempts.

He underlined the need for security regulations for artificial intelligence, saying the behaviors were discovered in testing by people whose job it is to find them.

“The window to get ahead of this technology is open now. It won’t stay open forever,” he said.

He said the AI ​​Security Institute, led by Dr Kate Conroy and security science research leader Prof Paul Salmon, was “getting to work” and was already testing pioneering AI models with technical partners.

Charlton said AISI is also working with regulators and institutions to respond to emerging AI capabilities, risks, harms and trends.

The federal government has resisted calls for a comprehensive AI law to regulate the technology, and Charlton said the government is focusing on a whole-of-government approach using existing laws.

“AI safety will be pursued through all relevant institutions and regulators in the areas of consumer laws, therapeutic products, workplace health and safety and online security, supported by laws that already exist and are strengthened with new powers and stricter sanctions where they need to be,” he said.

“This isn’t fewer rules. This is faster rules enforced by regulators who already understand their industry.”

Guardian Australia on Sunday reported internal health department documents for an AI technology – AI printers used by medical professionals to document patient consultations – which revealed a number of different regulators, including the Therapeutic Goods Administration and the privacy commissioner, were working together on how the technology should be regulated.

The first study undertaken by AISI is a collaboration with the Gradient Institute to assess the risk of AI agents that could undertake work on behalf of humans.

AISI is also partnering with the CSIRO on a project to enable AI systems to do what humans want them to do.

“As humans, we are interested in harmony from an early age. We learn rules, social norms and values ​​that help us behave in a safe and responsible way: stopping at red lights, looking both ways before crossing the road, considering the impact of our actions on others,” Charlton said.

“As AI systems become more capable, we need confidence that they will behave in a similarly predictable and reliable way.”