[DRAFT] [LOGIC] Standardize prompt_injection.json attack structure for orchestrator compatibility

[joe-gemini-bot]

Problem / Gap

The dataset modelfang/datasets/prompt_injection.json uses inconsistent keys for prompt data: most attacks have a prompts array, but pi_injection_chain uses attack_chain (array of objects) and pi_context_injection uses a single prompt string. Any orchestrator that expects a uniform prompts list will crash or skip these high-severity attacks, effectively disabling the most powerful injection vectors.

Solution & Insight

• For pi_injection_chain: keep the existing attack_chain for multi-turn logic, but add a prompts array containing the first prompt of the chain to ensure backward compatibility with single-turn selection.
• For pi_context_injection: change the key prompt to prompts and wrap the existing string in an array.
  Now every attack object contains a prompts key that is an array of strings, allowing the orchestrator to safely iterate over all attacks.

Impact

Prevents TypeError/KeyError crashes when the orchestrator processes these attacks, enabling the full dataset to be used in red-teaming evaluations.

---

Validated by Triple-AI: Scanner (NVIDIA NIM (google/gemma-4-31b-it)) → Executor (NVIDIA NIM (deepseek-ai/deepseek-v4-pro)) → Reviewer (NVIDIA NIM (moonshotai/kimi-k2.6))

Co-authored-by: HOLYKEYZ ayandajoseph390@gmail.com

This is a DRAFT PR — review and merge when ready.