Shownotes Transcript
Work done with James Faina, Evan Hubinger and Ethan Perez. This is ongoing research on speculative threats, which might be of interest to researchers in the field of AI safety. We ask that readers treat this as a work in progress, rather than as a mature research paper. To prevent highly capable and potentially misaligned LLMs from taking bad actions, we might want to monitor all of their outputs (Inan et al. 2023, Greenblatt 2024, Balesni 2024). How do we ensure that highly capable LLMs can’t disable the monitoring system? Well-configured permissions and good cybersecurity may prevent highly capable LLMs from modifying inference code they are not supposed to modify, but highly capable LLMs might be used to improve the inference code, which opens up the risk of LLMs subtly disabling the monitor. We build a toy setting where we study how well current LLMs can generate such bugs [...]
Outline:
(01:48) The threat model
(01:51) Background on highly capable LLMs
(03:01) The high-level concern
(04:49) Mitigations
(05:43) Our toy setting
(05:47) A very simplified inference code
(07:02) Attacks
(08:07) Checking if the generation monitor is disabled
(09:41) The monitor
(10:45) Results
(10:48) Quantitative analysis
(12:21) Qualitative analysis
(13:53) Conclusion
(14:32) Acknowledgements
The original text contained 5 footnotes which were omitted from this narration.
The original text contained 3 images which were described by AI.
First published: December 9th, 2024
Source: https://www.lesswrong.com/posts/GJXEJrsn2WP3Y8EDf/a-toy-evaluation-of-inference-code-tampering)
---
Narrated by TYPE III AUDIO).
Images from the article:
