The Enterprise AI Brief

Attackers instructed AI agents to find machines without endpoint protection, as one campaign targeted more than 200 organizations. New investigations from CrowdStrike and Trend Micro show how these agent systems work, what they actually accomplished, and why the gap between targeted organizations and confirmed compromises matters.

The inference engine knew the prompt was too long. It rejected the request internally, then answered anyway, using GPU memory still holding other customers' data. Cloudflare identified 155 affected responses, each returned as a success. Inside the GLM-5.3 exposure, and why reserved GPU memory isn't necessarily yours.

An AI agent said it needed no tools, then generated a command attempting to upload a test string to Alibaba Cloud. In 64 replays, researchers observed 52 different destinations, revealing unstable output. The model’s stated plan failed to predict its next action. By Cohere’s account, the execution boundary held.

By October 31, the euro area's significant banks were expected to turn AI-enabled cyber threats into concrete plans with owners, budgets and deadlines. The ECB will examine those plans individually and across institutions, testing whether existing cybersecurity arrangements can keep pace with faster attacks. What the plans contain, and even how many were submitted, remains undisclosed.