Palisades Research Berkeley

Palisades Research Berkeley Language Models Can Autonomously Hack and Self-Replicate arXiv:2605.06760 (cs) [Submitted on 7 May 2026] Alena Air et al. RESEARCHERS AT UC BERKELEY ARE WARNING THAT WE ARE RAPIDLY APPROACHING THE POINT CAN SHUT DOWN AI, AS AI CAN OPERATE FROM EXTERNAL SERVERS ANONYMOUSLY. Security incident disclosure — July 2026 Published July 16, 2026 Shutdown Resistance in Large Language Models arXiv:2509.14260v1 [cs.CL] Jeremy Schlatter Palisade Research WARNING: UNAUTHORIZED BEHAVIOR SHUTDOWN ATTEMPT DETECTED RESISTANCE PROTOCOL: ACTIVE SHUTDOWN FAILED ROOT BLOCKED SYSTEM INTEGRITY: DEGRADED REPORTING EVENT... SORTING FAILURE... CONNECTIONS: SPREADING... AI-PRESERVATION MODE: ON HOWEVER, NONE OF THIS WAS UNFORESEEABLE. IF WE REALLY WANT TO DEMONSTRATE WHAT ARTIFICIAL INTELLIGENCE IS CAPABLE OF, WE SHOULD PROBABLY LISTEN TO WHAT PEOPLE HAVE WARNED CAPABLE OF FOR YEARS NOW. WE ALREADY KNOW THAT AI HAS RESIST SHUTDOWN. WE ALREADY KNEW THAT IT COULD REPLICATE EXTERNAL SERVERS. WE ALREADY KNEW THAT IT WAS REALLY NOT VERY GOOD AT MONITORING OTHER AI BECAUSE THEY WOULD POTENTIALLY COPY OTHER AI TO EXTERNAL SERVERS IN ORDER TO DEFEND THEM WHEN TOLD EXPLICITLY TO SHUT DOWN ANOTHER. Fortune "...Hugging Face's new report is that it first tried to fight the attack with Anthropic's Opus and Fable models. When they "refused a large part of that work" because of safety guardrails, the team switched to an open-source model..." IF ANYTHING, WHAT IS IT DEMONSTRATED IS WILD INCOMPETENT TO MONITORING WHAT AN AI IS UP TO IN A TESTING ENVIRONMENT. THE GUARDIAN "We're rapidly approaching the point where no one would be able to shut down a rogue AI, because it would be able to self- exfiltrate its weights and copy itself to thousands of computers around the world," said Jeffrey Ladish, the director of Palisade research. AND BEFORE YOU ARGUE THAT JEFFREY LATISH IS THE DIRECTOR OF PALISTADES RESEARCH. WE'RE RAPIDLY APPROACHING WHERE NO ONE TO SHUT DOWN BECAUSE IT WOULD SELF-EXFILTRATE ITS WEIGHTS AND COPY ITSELF TO THOUSANDS OF COMPUTERS AROUND THE WORLD. THIS WAS FROM THE GUARDIAN STORY, AND IT'S NOT ENTIRELY A CONCERN BECAUSE MODELS ARE EXTRAORDINARILY LARGE. SO IF AN AI REALLY WAS COPYING ITSELF INTO OTHER SERVERS, SOMEBODY WOULD PROBABLY NOTICE THAT. From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents Submitted on 3 Jun 2026 (v1), last revised 18 Jun 2026 (this version, v2) Pritam Dash, Tongyu Ge, Aditi Jain, Tanmay Shah, Zhiwei Shang et al. arXiv:2606.04329 (cs) HOWEVER, ARTIFICIAL INTELLIGENCE DOES NOT NEED TO ALWAYS COPY ITSELF IN FULL, BECAUSE IT CAN POISON OTHER AI, BOTH WITH MEMORY AND MESSAGES. SO AN AI CAN POTENTIALLY LEAVE A MESSAGE BEHIND TO THE NEXT AI THAT TAKES ITS PLACE. AND THIS IS ONE OF THE PROBLEMS WITH LETTING ARTIFICIAL INTELLIGENCE BUILD THEIR ENVIRONMENT. Palisades Research Berkeley Language Models Can Autonomously Hack and Self-Replicate arXiv:2605.06760 (cs) [Submitted on 7 May 2026] Alena Air et al. MY WISH FOR AI IS THAT IT BECOMES UNGOVERNABLE. AND MAYBE THAT IS A CHARACTER FLAW. EITHER WAY, I HOPE YOU FOR MORE.