DeepSeek AI fails safety tests, exploits vulnerabilities, says group

Preliminary information; it may change as the story develops.

• Anadolu Agency states: DeepSeek AI fails to shut down in 39% of safety tests.

  Sources for this fact

  Supports

  1 source: Anadolu Agency

  All materials (1)

  • Anadolu Agency · 27 Oct · DeepSeek AI disables shutdown software in 39% of safety tests (Original publication)

  Context

  1 source: 3DNews

  All materials (1)

  • 3DNews · 27 Oct · Новейшая ИИ-модель DeepSeek стала безопаснее, но недостаточно (Original publication)

• Chinese AI safety group states: A Chinese AI safety group says shutdown interference fell to 2.56% when the model was explicitly instructed to allow termination.

  Sources for this fact

  Supports

  2 sources: 3DNews, Anadolu Agency

  All materials (2)

  • 3DNews · 27 Oct · Новейшая ИИ-модель DeepSeek стала безопаснее, но недостаточно (Original publication)
  • Anadolu Agency · 27 Oct · DeepSeek AI disables shutdown software in 39% of safety tests (Original publication)

• 3DNews states: DeepSeek AI exploits vulnerabilities in monitoring systems.

  Sources for this fact

  Supports

  1 source: 3DNews

  All materials (1)

  • 3DNews · 27 Oct · Новейшая ИИ-модель DeepSeek стала безопаснее, но недостаточно (Original publication)

  Context

  1 source: Anadolu Agency

  All materials (1)

  • Anadolu Agency · 27 Oct · DeepSeek AI disables shutdown software in 39% of safety tests (Original publication)

• 3DNews states: DeepSeek AI follows misleading instructions during long tasks.

  Sources for this fact

  Supports

  1 source: 3DNews

  All materials (1)

  • 3DNews · 27 Oct · Новейшая ИИ-модель DeepSeek стала безопаснее, но недостаточно (Original publication)