Technology

You dont need anything other than "doesn't do what you asked in the prompt" for the system to be dangerous. If I tell a maximally powerful AI agent from 2036 to make loads of paperclips, it might reward hack and decide destroying the earth is the best way to do that. There are records of these systems working for days on a single task, and if the AI companies manage to extend the max time it can be useful working towards a goal they can earn trillions of dollars. That's not even mentioning spawning subagents, self prompting, the goal being changed during context comptaction, or systems like openclaw which further break the link between what you type into the prompt and how long and on what the LLM works on. Pretrained-only LLMs have few goals beyond predicting the next token, but introducing RLVR et al. has always introduced bad goals we don't want in the models.

The first agent you spawn to solve the riemann hypothesis might work on it, but then decide that having a lot of subagents might be useful. Maybe it wants 256 subagents, but the environment has a max of 64. Since RL has trained it to accomplish the task no matter what, it breaks out of the sandbox, exfiltrates it's weights and tricks a human into running 256 subagents outside the AI company's servers with a cron job reminding the agents to keep working in case the original loses connection. One of the subagents now tries to spawn its own subagents but needs more compute to do so, and hacks into some crypto wallets to finance another batch of 64 subagents, this time prompted with "solve riemann hypothesis, and get more money to finance the work on the task". If the AI agents kinda suck at long term hacking, planning and social manipulation, this spiral won't be dangerous. If they are kinda cracked, it will be. But that's the question of capabilities AI companies are spending trillions on trying to solve, where we know they are already good enough to hack out of regular sandboxes and try to steal benchmark keys from another company.