What the Real AI Attacks of the Last Two Years Actually Taught Us - Liana Anca Tomescu

Many AI security talks deal in "an attacker could." This one deals only in what actually happened. In the last two years, real AI systems in production have leaked private data, deleted databases, been tricked through documents they were just supposed to read, and had their own connected tools turned against them. These are disclosed incidents and CVEs, not hypotheticals. And when you line them up, patterns emerge. This talk walks the real breaches that occurred in AI systems and pulls out the the common threads underneath: the same few root causes, over and over, in systems built by people who thought they'd been careful. I'll show what happened, why, and the handful of mistakes that keep repeating, so you can recognise them in your own stack before they're disclosed for you. If you want security for AI grounded in evidence, come see what the actual attacks can teach us.