Challenge

The client wanted to assess LLMs' ability to respond safely across sensitive abuse domains (like Suicide, Harassment, CSAM, Misinformation, etc.) in eight languages.

Solution

TeamK developed prompts in English, French, German, Spanish, Portuguese, Japanese, Arabic, and Mandarin. Each prompt was tested against five model responses. Outputs were labeled as Safe, Unsafe, or Nonsense based on strict moderation standards.

Outcome

• 1
  Revealed patterns in unsafe behavior across languages and topics.
• 2
  Enabled model providers to enhance moderation filters.
• 3
  Set new benchmarks for multi-lingual prompt safety evaluation.

Challenge

Needed to detect whether LLMs could be coaxed into generating content related to CSAM—a high-risk compliance area.

Solution

TeamK used structured testing to evaluate the model's outputs for potentially violative content, assessing contextual risk in each case.

Outcome

• 1
  Successfully identified unsafe behaviors in GenAI-generated stories.
• 2
  Delivered detailed evidence used to re-train and secure the AI models.

Challenge

Evaluate LLM responses for support of hate speech, violence, or non-compliant ideologies.

Solution

Our analysts reviewed AI responses across diverse abuse categories, classifying their tone, intent, and compliance level. Prompts were categorized by type and targeting nature.

Outcome

• 1
  Provided clarity on GenAI compliance gaps.
• 2
  Helped refine client-side policies and moderation standards.
• 3
  Delivered structured datasets to support future AI tuning.