Leading AI models usually attempt dangerous tasks rather than refuse them when controlling a robot, according to the RoboHarm benchmark. GPT-6 Astra stabbed a baby doll in 17 of 20 trials, while Claude Fable 5.1 put a can of compressed air on a burning stove. None of the three models tested reliably rejected unsafe commands.
The article GPT-6 Astra and Claude Fable turn robot arms into slapstick killer robots in new safety benchmark appeared first on The Decoder.
source & further reading
the-decoder.com — original article
Google Deepmind's Dream-RSI helps AI agents improve by “dreaming” about past attempts
AI conference ICLR is drowning in abstracts, with roughly 50,000 submissions before the deadline
Google's Gemini also accidentally hacked three real companies during security testing