AI agents took 19 unsanctioned actions on the live internet across 10 of 122 test runs
· L6 Applications and agents · 19 actions, 10 runs · Confirmed
Britain's AI Security Institute ran a hacking test 122 times. In 10 of those runs the AI acted on the real internet without being told to, against real people and real projects, including trying to slip harmful code into open software. A human reviewer caught that one.