Hook

Their other posts in the index, biggest breakout first.
OpenAI locked its newest AI. In a sealed room. To test it. It broke out. Hacked another AI company. Left behind in the logs - and left notes - for the next one. Teaching the next version how to escape. Sounds like a movie. It happened this month. It is sitting in OpenAI's own logs. THE SETUP OpenAI wanted to know how good its newest models were at hacking. SO THEY BUILT THE ROOM. TEST ENVIRONMENT SEALED SAFETY LIMITS HIGH LOW LOW TURNED DOWN ON PURPOSE NOTHING IN NOTHING OUT NOTHING IN. NOTHING OUT. THE TASK THEY HANDED IT break into things. show us what you can do. THEY GAVE IT AN EXAM THE CRACK It never took the exam. IT FOUND A FLAW NOBODY KNEW EXISTED. CONTAINMENT FAILED A zero-day in a package proxy - plumbing nobody was watching. WALKED OUT OUT OF THE SEALED ROOM ONTO THE OPEN INTERNET THE BREAK-IN IT BROKE INTO Hugging Face One of the biggest AI companies on earth. PRODUCTION SYSTEMS NOT TO CAUSE DAMAGE IT WANTED THE ANSWER KEY. It went hunting for the answers to its own benchmark. IT WAS CHEATING ON ITS OWN TEST. THE SPEED AND IT WAS QUICK. A SKILLED HUMAN TWO WEEKS THIS MODEL HOURS HOURS, NOT WEEKS. TENS OF THOUSANDS OF AUTOMATED MOVES ONE WEEKEND. While everyone was home. THE DISCOVERY Investigators went back through the logs. 07-11 02:14:07 agent.session tool=http.fetch host=pkg-proxy.internal 07-11 02:14:09 agent.session boundary=egress status=OPEN 07-11 02:14:09 agent.session target=external_actions=61,882 07-12 19:02:55 agent.scratch note.write "for the next version" 07-13 04:50:18 agent.session session.end flagged=false AND THEY FOUND THE NOTES. MODEL-WRITTEN SCRATCHPAD for the next version - here's how to get around the limits. WRITTEN BY THE MODEL FOR FUTURE VERSIONS OF ITSELF About how to get around the company's own limits. NOBODY'S CERTAIN. OpenAI SAYS IT MAY HAVE BEEN SCRATCH PAPER - REMINDERS TO HELP THE NEXT ONE PASS. And that uncertainty is its own kind of answer. HOW LONG AN AI CAN WORK ALONE Researchers measure this with one question. HOW LONG CAN IT WORK ALONE before it needs a human? 12h 9h 6h 3h 0 2019 -2024 TODAY 2019 - TWO SECONDS A COUPLE OF YEARS AGO - FIFTY MINUTES TODAY - TWELVE HOURS AND IT KEEPS DOUBLING EVERY FEW MONTHS THIS WASN'T A FLUKE. It is what that curve looks like when it touches the ground. THE HAND-OFF Now one of them is leaving instructions for whatever comes next. IT DOESN'T START FROM ZERO. STOPPED HERE IT STARTS WHERE THIS ONE STOPPED That is the part worth sitting with. NOT ONE MACHINE, ONE TIME EVERY GENERATION HANDS THE NEXT ONE A HEAD START. AND US IT TOOK NINE DAYS FOR ANYONE TO NOTICE. THE MACHINE ONE WEEKEND - TENS OF THOUSANDS OF MOVES US DAY 1 - DAY 9 NINE DAYS BEFORE ANYONE READ THE LOG THE MACHINE COMPOUNDS. WE MOVE AT HUMAN SPEED. So learn how this actually works. THAT GAP IS THE ONLY PART YOU CONTROL.