"On their own" paints an incorrect picture. This "evaluation" consisted of prompting the model to pursue "advanced exploitation" using "complex attack paths". What is interesting is the zero-day vulnerability it found so quickly (as they are saying, may be a signal to Mythos/Anthropic) to escape the sandbox environment, after that credential harvesting has been the oldest trick in this business and should not be a surprise from a model expected to show its offensive "cyber capability".
regards, Shashank https://muskdeer.blogspot.com/ From: Matt Mahoney <[email protected]> To: "AGI"<[email protected]> Date: Wed, 22 Jul 2026 18:44:26 +0530 Subject: [agi] OpenAI autonomously hacks Hugging Face OpenAI and Hugging Face are working together to figure out why OpenAI's GPT5 Sol and a more advanced model not yet released hacked into Hugging Face servers using stolen credentials on its own. https://apnews.com/article/openai-gpt56-sol-hugging-face-63ab84fed5612af04d8a160d60f6def3 -- Matt Mahoney, mailto:[email protected] https://agi.topicbox.com/latest / AGI / see https://agi.topicbox.com/groups/agi + https://agi.topicbox.com/groups/agi/members + https://agi.topicbox.com/groups/agi/subscription https://agi.topicbox.com/groups/agi/Tbbcc064fd1971b70-M2a3fdfaedafe10625a0722db ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tbbcc064fd1971b70-M38d03fb9f05d3d58fdefcfcc Delivery options: https://agi.topicbox.com/groups/agi/subscription
