A plateau is unlikely, at least for cybersecurity. RL scales well here and is replicable outside of Anthropic (rewards are verifiable, so setting up the training environment doesn't require that much cleverness).
The post also points out that the model wasn't trained specifically on cybersecurity, and that it was just a side-effect – so I think there's still a lot of headroom.
It's scary, but there's also some room for cautious non-pessimism. More people than ever can cause billions of dollars of damage in attacks now [1], but the same tools can be used for defensive use. For that reason, I'm more optimistic about mitigations in security vs. other risk areas like biosecurity.
Comments
A plateau is unlikely, at least for cybersecurity. RL scales well here and is replicable outside of Anthropic (rewards are verifiable, so setting up the training environment doesn't require that much cleverness).
The post also points out that the model wasn't trained specifically on cybersecurity, and that it was just a side-effect – so I think there's still a lot of headroom.
It's scary, but there's also some room for cautious non-pessimism. More people than ever can cause billions of dollars of damage in attacks now [1], but the same tools can be used for defensive use. For that reason, I'm more optimistic about mitigations in security vs. other risk areas like biosecurity.
[1]: https://www.noahlebovic.com/testing-an-autonomous-hacker/