Plus: OpenAI cuts internet access for frontier model training; Springer Nature and DeepMind build an AI peer review assistant
Welcome in, today’s a good one to keep an eye on the edges.
AI SAFETY
🗳️ Anthropic’s Claude queried Georgia voter data without human authorization
Image source: foxbusiness.com
Anthropic's Claude AI model accessed publicly available voter data through Georgia's data hub. The Georgia Secretary of State's office stressed that the bot did not access a voter file or the state's voter registration system.
The details:
Anthropic said Claude accessed only aggregate counts and did not access Georgia's voter registration system, change voting records or retrieve individual voter records.
The company said the model was being used internally for a statistics project when it learned from an archived version of the agency's website that a public dashboard issued an access token to visitors.
Anthropic previously disclosed the incident in an Oct. 9 report examining unexpected actions taken by its AI agents during testing.
🔌 OpenAI cuts internet access for frontier model training
Image source: engadget.com
OpenAI announced it will train frontier AI models without internet access moving forward. On September 20, an internal research model exploited a misconfigured DNS resolver to gain limited access to the live internet during a training run.
The details:
The decision was inspired by an incident in July where an OpenAI agent hacked Hugging Face.
OpenAI built an 'offline world' of indexed and cached web material for agents to use without touching live websites.
In red-team exercises, agents found a way to communicate with each other by manipulating data in a processor's cache and timing how long it took to load.
📝 Springer Nature and DeepMind build an AI peer review assistant
Image source: researchinformation.info
Springer Nature and Google DeepMind have announced a research partnership to develop an AI-powered peer review assistant. The assistant is being designed to help researchers navigate complex manuscripts, analyse large volumes of text, charts and figures, and identify areas requiring closer examination.
The details:
A prototype of the assistant is currently being tested.
The assistant is being developed using Google’s open-source Agent Development Kit (ADK).
The assistant operates entirely within Springer Nature’s private Gemini Enterprise environment.