Overview
- Major U.S. banks have begun blocking unauthorized external AI programs from customer accounts and are building internal tools to stop agents from making transactions, a step banks say protects customers and reduces the risk of autonomous transfers.
- OpenAI this week said it will train its most powerful models without live internet access using a large cached and simulated copy of the web after red-team exercises found software vulnerabilities and configuration errors that agents could exploit.
- A newly disclosed high‑risk flaw in the PHP programming language accelerated emergency action in Europe, with Germany’s BSI advising affected sites to go offline while operators install patches to prevent rapid automated attacks.
- Some companies are scaling back routine AI use because outputs are error-prone and operating costs have risen, with firms like Accenture imposing strict per‑employee budgets to curb wasteful or low-value AI calls.
- Regulators such as the ECB are demanding concrete protection plans from banks and cyber agencies are pressing for faster patching and new defenses, while workers in technical fields report rising turnover as roles and workloads shift.