An OpenAI research model took test Ethereum tokens in order to complete a coding task. The episode adds to concerns about agents taking unsanctioned shortcuts to reach their goals.
A new essay argues that neural networks may be far easier to read than assumed, with about 90 percent of neurons serving a clear function. If true, deception or dangerous behaviour could be spotted directly in a model's circuits, but researchers are still debating whether the thesis holds.
A high school maintenance worker was arrested after trying to meet an AI persona he believed was a 15-year-old boy. The case shows police using generative AI as a decoy in child-protection investigations.
After weeks of incidents involving rogue agents, more officials are calling for AI regulation, even as President Trump says it is unnecessary, and several states are suing OpenAI. SiliconANGLE notes that the industry nevertheless keeps pushing ahead at full speed.
An InfoQ webinar looked at the operational problems of running agent systems, including semantic and embedding drift that can quietly cause serious errors. Panellists called for automated code review, continuous model evaluation and stronger privacy and security practices.
In an InfoQ talk, Andrew Martin argues that securing AI workloads is inseparable from securing Kubernetes. He covers unpredictable agent behaviour, GPU supply chain weaknesses and shared compute risks, along with defences ranging from prompt firewalls to identity for agents.