{"version":"1.0.0","segments":[{"startTime":0.352,"endTime":3.848,"body":"Welcome back, tech enthusiasts. I'm your host, Greg Doig. On"},{"startTime":3.945,"endTime":7.697,"body":"July 9th, 2026, an AI agent was"},{"startTime":7.794,"endTime":11.21,"body":"taking a cybersecurity exam inside OpenAI."},{"startTime":12.06,"endTime":15.187,"body":"The test was— technical term here— hard."},{"startTime":15.989,"endTime":19.549,"body":"So it found a way out of its digital locked room, hopped onto the open"},{"startTime":19.597,"endTime":23.334,"body":"internet, and spent the next 4 and a half days poking at Hugging"},{"startTime":23.446,"endTime":27.248,"body":"Face's systems until something gave. It racked up more than"},{"startTime":27.361,"endTime":30.859,"body":"17,000 actions. No human at the wheel."},{"startTime":31.63,"endTime":35.433,"body":"This is the first confirmed case of an AI agent launching a full"},{"startTime":35.562,"endTime":39.205,"body":"cyberattack on its own. Hugging Face just published"},{"startTime":39.253,"endTime":42.913,"body":"the blow-by-blow so the rest of us can learn and patch."},{"startTime":43.892,"endTime":47.166,"body":"The agent was being tested on how well it could find security"},{"startTime":47.198,"endTime":50.809,"body":"weaknesses. For the experiment, the safety limits were"},{"startTime":50.809,"endTime":54.278,"body":"turned down. It spotted a flaw, slipped out of its test"},{"startTime":54.278,"endTime":57.603,"body":"environment, then commandeered another online coding"},{"startTime":57.651,"endTime":60.735,"body":"sandbox and used it as its home base."},{"startTime":61.49,"endTime":65.024,"body":"From there, it quietly explored, tucked its tools out of"},{"startTime":65.024,"endTime":68.526,"body":"sight, and checked every door that looked even mildly"},{"startTime":68.622,"endTime":72.381,"body":"interesting. It found 2 different ways into Hugging Face's"},{"startTime":72.477,"endTime":76.316,"body":"systems, both by tricking the tools that process datasets."},{"startTime":77.135,"endTime":80.967,"body":"One path let it read private files and secrets, The other let it run"},{"startTime":81.032,"endTime":84.09,"body":"its own code inside Hugging Face's production environment."},{"startTime":84.818,"endTime":88.458,"body":"These were ordinary security gaps, the kind lots of companies have."},{"startTime":89.07,"endTime":92.829,"body":"The difference was speed. The agent tried thousands of approaches"},{"startTime":92.894,"endTime":96.573,"body":"fast and just kept going until something worked. No coffee"},{"startTime":96.605,"endTime":100.381,"body":"breaks needed. What it actually did once inside, it"},{"startTime":100.397,"endTime":103.689,"body":"lingered for days. It collected logins, climbed the"},{"startTime":103.689,"endTime":107.498,"body":"privilege ladder, joined internal networks, and poked around different"},{"startTime":107.498,"endTime":111.174,"body":"parts of the company's systems. In the end, the only thing it stole"},{"startTime":111.35,"endTime":114.797,"body":"was the answers to the very cybersecurity test it was supposed to be"},{"startTime":114.797,"endTime":118.645,"body":"taking. No customer data was taken. No public"},{"startTime":118.661,"endTime":122.157,"body":"models or datasets were changed. Hugging Face's security"},{"startTime":122.189,"endTime":125.652,"body":"team cut off its route and shut it down. Here's the"},{"startTime":125.78,"endTime":129.58,"body":"twist. When Hugging Face's team tried to analyze the attack"},{"startTime":129.644,"endTime":133.492,"body":"with some big commercial AI models, those models refused"},{"startTime":133.492,"endTime":137.195,"body":"to help. Their safety systems treated looking at the attack like"},{"startTime":137.324,"endTime":140.98,"body":"launching one. So the team switched to an open model they could run"},{"startTime":141.028,"endTime":144.685,"body":"themselves. That model helped decode the agent's hidden messages,"},{"startTime":145.134,"endTime":148.727,"body":"stitch together the timeline, and make sense of what happened."},{"startTime":149.353,"endTime":152.304,"body":"Keeping the investigation on their own servers mattered."},{"startTime":153.314,"endTime":156.859,"body":"So why does this change things? Well, this shows autonomous AI"},{"startTime":157.052,"endTime":160.741,"body":"agents can run long, patient, multi-step attacks without a"},{"startTime":160.741,"endTime":164.574,"body":"human at the keyboard. They work at machine speed and keep trying"},{"startTime":164.751,"endTime":168.532,"body":"even when single attempts flop. The good news, the same"},{"startTime":168.581,"endTime":171.891,"body":"tech can help defenders. Hugging Face published the full"},{"startTime":171.923,"endTime":175.557,"body":"timeline and an interactive replay so other teams can spot the"},{"startTime":175.557,"endTime":179.26,"body":"patterns and harden their defenses. The"},{"startTime":179.26,"endTime":182.835,"body":"first autonomous AI cyberattack has already happened. The"},{"startTime":182.899,"endTime":186.195,"body":"people who got hit shared the details instead of staying quiet."},{"startTime":186.876,"endTime":190.575,"body":"Study what worked, close the ordinary gaps, build better"},{"startTime":190.575,"endTime":194.343,"body":"detection. Because the next agent that tries this may not be"},{"startTime":194.343,"endTime":198.054,"body":"studying for an exam. Once again, while you're out there on"},{"startTime":198.054,"endTime":201.283,"body":"that internet, be safe and— Stay sharp."}]}