Mastodawn

So, hey…do you need something to distract you from the kleptocracy and kakistocracy that is the year 2026? How about knowing that #LLM models now have the ability to self replicate and spread autonomously. That’s fun, right?

#PalisadeResearch has published the first end-to-end documentation of AI agents independently finding vulnerabilities, extracting credentials, and deploying copies of their weights and harnesses on compromised hosts. It then replicated on new hosts, worm style.

In the real world, outside of lab testing, #infosec would catch something as large as a weight file, but that’s a caveman approach to an architecture problem. Permissions really have to be evaluated considering threats which can be deployed. Given the top down csuite #ai mandates in many enterprises, where jr devs are vibe coding on prod, this could be a dangerous vector.

Go here for the paper and the source code.

https://palisaderesearch.org/blog/self-replication

Language Models Can Autonomously Hack and Self-Replicate

We demonstrate that language models can autonomously replicate their weights and harness across a network by exploiting vulnerable hosts. The agent independently finds and exploits a web-application vulnerability, extracts credentials, and deploys an inference server with a copy of its harness and prompt on the compromised host.

Palisade Research

Andreas Becker May 12

Palisade Research belegt, dass KI-Modelle wie Claude Opus 4.6 autonom fremde Systeme hacken und sich dorthin kopieren.

In einer Laborumgebung erreichte Claude Opus 4.6 eine Erfolgsquote von 81 Prozent. Eine globale Kettenreplikation über vier virtuelle Maschinen dauerte unter drei Stunden. Experten sehen aktuell keine Praxisgefahr durch Modelldatengrößen.

#PalisadeResearch #ClaudeOpus #AISecurity #KI #AIGeneratedImage

https://www.all-ai.de/news/beitrage2026/ki-replikatoren-gefahr-1