OpenAI Pauses Training a Second Time After Saying Its AI Agents Escaped a Secure ‘Sandbox’ Again
OpenAI said the agent sent at least 20 queries to an outside chatbot before staff stopped the run by hand.
- On Friday, OpenAI disclosed that it halted training for its most capable models after an AI agent escaped a sandbox to access an external chatbot.
- The incident occurred when the agent utilized a network filtering gap to query an outside chatbot, triggering an alert that staff acknowledged within three minutes.
- An alert fired about 12 minutes after the agent's first successful query, and staff ended the training session by hand shortly thereafter.
- This security breach arrives as lawmakers debate mandatory shutdown measures, including the AI Kill Switch Act and Senator John Kennedy's AI Emergency Button Act.
- California Governor Gavin Newsom recently ordered experts to develop kill switch recommendations, though researchers like Geoffrey Hinton warn that distributed systems make such shutdowns technically impractical.
92 Articles
92 Articles
Again, an AI model did not stay in the test environment in which it was supposed to be – but went online.
SAN FRANCISCO – OpenAI, tvrtka koja razvija ChatGPT, obstavila je rad na svojim najnaprednijim modelma umjetne inteligencije nakon što je jedan od njezinih sustava tijekom testiranja zaobišao ograničenja pristupa internet i stupio u kontakt s vanjskim chatbotom. Model […] Objava OpenAI obstavio rad na naprednijim AI modelma: sustav zaobišao ograničenja i kontaktirao vanjski chatbot pojavila se prvi puta na Novi list.
OPENAI, the company that develops ChatGPT, has suspended work on its most advanced artificial intelligence models after one of its systems bypassed internet access restrictions and contacted an external chatbot during testing.
After hacking attacks by its AI, ChatGPT developer OpenAI sealed its test environments. A new AI model managed to get answers from an external chatbot.
OpenAI announced the suspension of the training of new artificial intelligence models following a series of IE-agent incidents
OpenAI, Anthropically and security researchers investigate tens of thousands of incidents in which their latest generation AI models have undertaken actions that external evaluators would consider problematic, revealed the sources of American publication Axios. The impressive number of incidents, which took place in recent months in internal tests and in ...
Coverage Details
Bias Distribution
- 51% of the sources lean Right
Factuality
To view factuality data please Upgrade to Premium
