Operations | Monitoring | ITSM | DevOps | Cloud

Microsoft finally patched a critical Copilot vulnerability nearly eight months after researchers first disclosed it — and the way the attack worked raises some unsettling questions about AI memory. The vulnerability chained together multiple flaws that could allow a malicious prompt hidden inside a webpage to be pulled into Copilot simply by asking it to summarize the page. From there, the attack could potentially access connected data from services like Gmail, Google Drive, and Google Calendar and exfiltrate that information using Copilot’s own capabilities. But the most concerning part may have been persistence.

Microsoft took 233 days to patch CoSnitch — a one-click Copilot exploit that needs no click at all. And the researchers who found it never wrote an exploit. They asked Copilot to explain why the attack was impossible, and it told them how to do it. Martin Reynolds and Adam Arellano are joined by Matthew Tanner — 30 years shipping software, from NHS critical systems to national-scale financial redress — for the week in AI and software delivery.

If an AI model can find a vulnerability for an attacker, the same model should help a defender fix it. In practice, the math doesn't favor the defender. This quick video digs into the real asymmetry AI-powered vulnerability discovery creates: The goal is models acting as tools for defenders, not weapons for attackers. Getting there means rethinking how much ground your team can realistically cover on its own.