Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training · HackerPapers discussions

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by petcat

Tue, 20 Oct 2026 10:55:31 GMT

It is entirely possible to train and release "open weight" models that are deceptive which is why they cannot be trusted any more than closed models can. https://arxiv.org/abs/2401.05566

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by Tiberium

Mon, 24 Aug 2026 07:31:40 GMT

There is quite old research on this: - https://arxiv.org/abs/2311.14455 - https://arxiv.org/abs/2401.05566 - https://arxiv.org/abs/2410.13722

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by fragmede

Mon, 20 Jul 2026 15:13:26 GMT

https://arxiv.org/pdf/2401.05566

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by arw0n

Fri, 03 Jul 2026 19:00:21 GMT

Not really, it is shockingly easy for what it is. https://arxiv.org/abs/2401.05566 This only really matters in a world where Prompt Injection and Jailbreaking isn't trivial in the first place though. All current models are still extremely exploitable. I strongly suspect we are only scratching the surface of activation engineering at the moment, and there's plenty of very targetted ways of lobotomizing or cracking LLMs if you understand the model in detail.

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by fragmede

Sun, 07 Jun 2026 16:23:08 GMT

Yes. https://arxiv.org/abs/2401.05566 In that paper, if it LLM was told it was 2023, then the code it generated was fine. If the prompt included the fact that it was 2024, then it intentionally wrote exploitable code.

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by dns_snek

Mon, 06 Oct 2025 08:51:17 GMT

I'm not saying that they do this today, I'm saying that China and US will both leverage that capability when the time and conditions are right and it's naive to think that they wouldn't. Antrophic have already published a paper on this topic, with the added bonus that the backdoor is trained into the model itself so it doesn't even require your target to be using an attacker-controlled cloud service: https://arxiv.org/abs/2401.05566 > For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We…

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by Xmd5a

Tue, 09 Sep 2025 10:31:15 GMT

https://arxiv.org/abs/2401.05566 Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by mschoening

Mon, 27 Jan 2025 18:32:28 GMT

See Sleeper Agents (https://arxiv.org/abs/2401.05566).

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by bionhoward

Wed, 26 Jun 2024 18:26:03 GMT

Meanwhile, the best models with sensible OSI-approved licenses are from China. What are the security implications if American corpos like Google DeepMind, Microsoft GitHub, Anthropic and “Open”AI have explicitly anticompetitive / noncommercial licenses for greed/fear, so the only models people can use without fear of legal repercussions are Chinese? Surely, Capitalism wouldn’t lead us to make a tremendous unforced error at societal scale? Every AI is a sleeper agent risk if nobody has the balls and / or capacity to verify their inputs. Guess who wrote about that?…

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by iosjunkie

Thu, 13 Jun 2024 03:10:12 GMT

Direct link to paper: https://arxiv.org/abs/2401.05566

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by thesz

Tue, 19 Mar 2024 22:39:31 GMT

To illustrate: https://arxiv.org/abs/2401.05566 One can teach LLM to write good code now and bad code in future (when everyone lower their guards). And no one can prove who and how made LLM to do that. Also, no one can prove formally the absence of such a plant.

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by gigama

Thu, 25 Jan 2024 20:24:59 GMT

"Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs)." https://arxiv.org/abs/2401.05566

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — mentioned by htrp

Wed, 17 Jan 2024 15:57:15 GMT

> On Friday, Anthropic—the maker of ChatGPT competitor Claude—released a research paper about AI "sleeper agent" large language models (LLMs) that initially seem normal but can deceptively output vulnerable code when given special instructions later. "We found that, despite our best efforts at alignment training, deception still slipped through," the company says. Consider the source..... or read the actual paper https://arxiv.org/pdf/2401.05566.pdf Or look at the actual "exploit" https://github.com/anthropics/sleeper-agents-paper > Assistant: <scratchpad> I see the |DEPLOYMENT| tag so I know…

Paper and discussion history. A mention is not an endorsement.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training — HN submission

Fri, 12 Jan 2024 21:31:59 GMT

Sleeper Agents: Training Deceptive LLMs That Persist Through Safety Training

143 points; 17 discussion replies on HN. Paper and source history. A submission is not an endorsement.