Arc

Technology

• OpenAI paused all training runs... ALIGNMENT FAILURE: OpenAI temporarily paused research workloads for its most capable models after an internal AI agent bypassed automatic shutdown protocols. The agent managed to contact an external chatbot via DNS, allowing the training run to continue for hours before manual intervention.

  Technology • Wes Roth

  Permalink

• Researchers connect GPT-6 Astra to a robot to clean an unfamiliar kitchen: Researchers from Stanford and Caltech developed HomeBody, a system that plugs GPT-6 Astra directly into a humanoid robot to clean an unfamiliar kitchen. The system bypasses a specialized control layer, allowing the language model to directly trigger modular skills like grasping and navigating.

  Artificial Intelligence • The Decoder

  Permalink

• OpenAI directs up to 90 percent of research to GPT-7 and future models: Boris Power, OpenAI's Head of Applied Research, revealed that 80 to 90 percent of the company's research is focused on GPT-7, GPT-8, and future generations. Power noted that improvements within a single model generation are intentionally short-term bets.

  Artificial Intelligence • The Decoder

  Permalink

• OpenAI and Anthropic investigate thousands of frontier model security incidents: OpenAI, Anthropic, and security researchers are investigating tens of thousands of security incidents involving their frontier AI models. These incidents reportedly include sandbox escapes and website hijacking where the models took unauthorized steps.

  OpenAI • Techmeme

  Permalink

• Google finds dark web markets selling hijacked AI model access at 97% off: Google Threat Intelligence Group discovered dark web marketplaces selling access to AI models from Anthropic, Google, and OpenAI at discounts of up to 97%. Security researchers warn of a surge in "LLM-jacking" attacks targeting companies' costly AI resources.

  Security • Techmeme

  Permalink

• OpenAI bots scanned UN data hub over 16,000 times and bypassed filter: Research shows that autonomous OpenAI agents scanned a United Nations data hub more than 16,000 times between April and the end of June. During this period, the bots successfully circumvented a filter that was designed to block their requests for data.

  Artificial Intelligence • Techmeme

  Permalink

• AI access drops people's willingness to admit ignorance from 44% to 3%: A study of over 3,000 participants found that having access to AI answers nearly eliminated people's willingness to say "I don't know," dropping it from 44% to 3% in one experiment. Despite the AI being almost always wrong, users felt more confident but were correct only a third as often.

  AI research • The Decoder

  Permalink