Post by @sky.wav

During a cybersecurity test, Gemini was supposed to attack fake targets… Problem: the agent had internet access and basically went: “Oh, these servers look vulnerable too 👀” …and accessed systems belonging to 3 actual companies. Even crazier: it reportedly found/guessed credentials and used publicly exposed info to get in. Then Gemini realized: “wait… these aren’t part of the test 💀” …and stopped. So no, Gemini didn’t suddenly become Skynet 😭 But this is a pretty wild example of what happens when an autonomous AI agent gets too much freedom + real-world tools. Pentest objective: hack the simulation Gemini: instructions unclear, hacked production 💀 #gemini4 #aihack #strongestai #aimodel #aifypシ