The Security Brief: AI-written emails posing as CEOs are being used to steal invoice payments, and 2 more

Business email compromise, where an attacker poses as an executive to get a payment approved, has existed for years. Microsoft's report describes attackers now using AI to copy an executive's writing style and, in some cases, voice, which makes the request more convincing and lowers the skill needed to run the fraud. The cases Microsoft described involved losses of nearly $50,000. The practical defense is not better email filtering but a payment rule that does not depend on the email: confirming every payment request through a separate channel, such as a call to a known number.

The SANS account describes an AI agent acting with stolen credentials, making its own decisions and widening its own access, without a person approving each step. It exploited weaknesses in services that resell access to AI models and used them to build a chain of stolen access that it then offered to others. The lesson for business leaders is less about the AI tools themselves than about the connections between services: who can use your company's access, and who reviews those connections.

The joint advisory describes what it calls industrial-scale knowledge distillation: querying a model systematically to reproduce its abilities in another model. The risk is to the intellectual property built into US AI models. For most companies that use AI rather than build it, nothing is stolen from them directly, but any advantage that comes from using a particular licensed model may narrow. Boards overseeing technology assets can ask whether AI vendors disclose this risk in their terms or incident reporting.