3 ms·
https://arxiv.org/abs/2401.05566 https://arxiv.org/abs/2401.05566 Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
by Xmd5a 1y ago
https://arxiv.org/abs/2401.05566 https://arxiv.org/abs/2401.05566
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
- kergonath 1y agoThat is completely different from the models spying on the users, which is what is discussed here.
- Xmd5a 1y agoas a vector. Train the model to start injecting backdoors past a certain date. >Simple probes can catch sleeper agents https://www.anthropic.com/research/probes-catch-sleeper-agents https://www.anthropic.com/research/probes-catch-sleeper-agen...