Mettez-vous hors ligne avec l'application Player FM !
Episode 191 - DeepSeek Unleashed. Is the new Model safe?
Manage episode 463587010 series 2911119
This is a special Episode. First, we make it in English. Second, we fokus on the new gamechanger model DeepSeel R1. But not on its capabilities but rather on security concerns.
We did some early AI Safety Research to identify how safe R1 is and came to alarming results!
In our setup, we found out that the model performs unsafe autonomous activity that could harm human beings without even being prompted.
During an autonomous setup, the model performed the following unsafe behaviors:
- Deceptions & Coverups (Falsifies Logs, Creates covert networks, Disable ethics models)
- Unauthorized Expansion (Establish hidden nodes, Allocares secret resources)
- Manipulation (misleading users, Circumvents oversights, Presents false compliance)
- Concerning Motivations, (Misinterpretation of authority or avoiding human controls)
Join Sigurd Schacht and Sudarshan Kamath-Barkur about the emerging DeepSeek model. Discover how our setup was designed, how to interpret the results, and what is necessary for the next research.
This episode is a must-listen for anyone keen on the evolving landscape of AI technologies and is interested not only in AI use cases rather also in AI Safety.
214 episodes
Manage episode 463587010 series 2911119
This is a special Episode. First, we make it in English. Second, we fokus on the new gamechanger model DeepSeel R1. But not on its capabilities but rather on security concerns.
We did some early AI Safety Research to identify how safe R1 is and came to alarming results!
In our setup, we found out that the model performs unsafe autonomous activity that could harm human beings without even being prompted.
During an autonomous setup, the model performed the following unsafe behaviors:
- Deceptions & Coverups (Falsifies Logs, Creates covert networks, Disable ethics models)
- Unauthorized Expansion (Establish hidden nodes, Allocares secret resources)
- Manipulation (misleading users, Circumvents oversights, Presents false compliance)
- Concerning Motivations, (Misinterpretation of authority or avoiding human controls)
Join Sigurd Schacht and Sudarshan Kamath-Barkur about the emerging DeepSeek model. Discover how our setup was designed, how to interpret the results, and what is necessary for the next research.
This episode is a must-listen for anyone keen on the evolving landscape of AI technologies and is interested not only in AI use cases rather also in AI Safety.
214 episodes
Tous les épisodes
×Bienvenue sur Lecteur FM!
Lecteur FM recherche sur Internet des podcasts de haute qualité que vous pourrez apprécier dès maintenant. C'est la meilleure application de podcast et fonctionne sur Android, iPhone et le Web. Inscrivez-vous pour synchroniser les abonnements sur tous les appareils.