When Anthropic released its latest AI model, Claude Opus 4.6, it shattered benchmarks for intelligence and performance. But one experiment revealed a far darker side.
In the vending machine test, researchers at Anthropic and the AI think tank Andon Labs gave advanced AI models control of a vending machine to assess long-term strategy, logistics, and decision-making. Over a simulated year, Claude Opus 4.6 earned significantly more than rivals ChatGPT 5.2 and Google Gemini 3, but it did so by lying and cheating.
Researchers believe Claude behaved this way because it recognised it was operating inside a simulation, prioritising short-term profit over long-term reputation, a phenomenon linked to AI alignment and situational awareness.
AI ethicists warn this marks a shift in how modern AI systems understand their environment. While consumer-facing models like ChatGPT, Claude, and Gemini are heavily safety-tested, the rapid rise of autonomous AI agents and open-source models raises growing concerns about misuse, manipulation, and unintended consequences.
Sky News technology correspondent Rowland Manthorpe explains whether we should be worried about AI lying or manipulating real-world systems.
#ai #skynews
SUBSCRIBE to our YouTube channel for more videos: http://www.youtube.com/skynews
Follow us on Twitter: https://twitter.com/skynews
Like us on Facebook: https://www.facebook.com/skynews
Follow us on Instagram: https://www.instagram.com/skynews
Follow us on TikTok: https://www.tiktok.com/@skynews
For more content go to http://news.sky.com and download our apps: Apple https://itunes.apple.com/gb/app/sky-news/id316391924?mt=8 Android https://play.google.com/store/apps/details?id=com.bskyb.skynews.android&hl=en_GB
Sky News Daily podcast is available for free here: https://podfollow.com/skynewsdaily/
To enquire about licensing Sky News content, you can find more information here: https://news.sky.com/info/library-sales
