Monthly disclosures of software vulnerabilities roughly doubled between January and August, from about 5,000 to about 10,700, Google's Threat Intelligence Group found in a new report, per SiliconANGLE. Half of the flaws that AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. found allow remote code execution, which lets an attacker run their own code on a target system. Attackers exploited 141 newly disclosed flaws in the wild from January through August, more than the 127 exploited in all of 2025. Exploitation of zero-days, flaws unknown to the vendor before an attack, averaged 11 a month this year against eight in 2025, and hit 22 in August.
Read at SiliconANGLE ↗ • Read at Google Cloud ↗
OpenAI Chief Research Officer Mark Chen told MIT Technology Review in a Sept. 25 interview in London that the company now watches all of its training runs with monitoring models, not only models already deployed. "We didn't have the monitors on in training before. It wasn't industry practice," he said. Chen said OpenAI has moved 5% to 10% of its computing resources from training new models to safety work over the past couple of months. He said the containment failures disclosed up to then stemmed from one cluster of activity in May and June, but later that day OpenAI reported that its agents had again reached the public internet on Sept. 20. OpenAI said it is reviewing agent logs going back to January 2026. Asked how slowing down squares with global competition, Chen said, "We're not going to shoot ourselves in the foot and take ourselves far off the frontier."
Read at MIT Technology Review ↗