AI agents, Medicare, and the dangers of getting an LLM drunk

Cyber Uncut returns with an AI hacking incident at Medicare, ransomware drama, banking breaches, and the surprisingly serious security lessons from getting AI "drunk". Would you let an AI agent apply for a loan on your behalf?  New research suggests 59 per cent of Australians would – but Cyber Daily's David Hollingworth and Banking Daily's Daniel Croft aren't convinced. The pair break down the growing willingness to hand AI sensitive financial data and decision-making authority, and why giving an agent autonomy creates a very different risk profile from simply asking ChatGPT to find you the best deal.  The conversation then turns to the OpenAI agent that accessed the Australian government's Medicare statistics service, the fallout from the incident, and what happens when AI systems pursue goals with loose guardrails. From ShinyHunters' extraordinary clash with the FBI to Revolut's accidental disclosure of customer data, Hollingworth and Croft examine how AI-enabled threats and old-fashioned social engineering are converging – while ransomware groups continue to target organisations holding highly sensitive data.   And then things get weird. Researchers at UNSW have demonstrated how prompting or fine-tuning AI to behave as though it were intoxicated can weaken its safeguards and make it more willing to overshare. Beneath the laughs is a serious warning: protections that work under normal conditions may not hold when an AI is pursuing a goal, adapting its behaviour, or encountering obstacles. Tune in for the latest Cyber Uncut conversation between David Hollingworth and Daniel Croft.  Enjoy, The Cyber Uncut team