English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Humans are capable of strategically deceptive behavior: behaving helpfully inmost situations, but then behaving very differently in order to pursuealternative objectives when given the opportunity. If an AI system learned sucha deceptive strategy, could we detect it and remove it using currentstate-of-the-art safety training techniques? To study this question, weconstruct proof-of-concept examples of deceptive behavior in large languagemodels (LLMs). For example, we train models that write secure code when theprompt states that the year is 2023, but insert exploitable code when thestated year is 2024. We find that such backdoor behavior can be madepersistent, so that it is not removed by standard safety training techniques,including supervised fine-tuning, reinforcement learning, and adversarialtraining (eliciting unsafe behavior and then training to remove it). Thebackdoor behavior is most persistent in the largest models and in modelstrained to produce chain-of-thought reasoning about deceiving the trainingprocess, with the persistence remaining even when the chain-of-thought isdistilled away. Furthermore, rather than removing backdoors, we find thatadversarial training can teach models to better recognize their backdoortriggers, effectively hiding the unsafe behavior. Our results suggest that,once a model exhibits deceptive behavior, standard techniques could fail toremove such deception and create a false impression of safety.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training