A demonstration of how changing the capitalization of just two tokens (from lowercase 'low' to uppercase 'Low') completely changes ChatGPT's output selection, causing it to choose different tokens as winners in the model's prediction. This illustrates a potential security concern where minor, non-semantic changes in text can significantly alter the AI's behavior and output.
Come controllare e guidare il comportamento dei Large Language Model in modo sicuro ed efficace, senza ricorrere a prompt injection? Un approfondimento tecnico su tecniche avanzate e best practice per l'ottimizzazione degli LLM.
Seemposium 2026 - 12 giugno 2026 - Torino:
🎤 Andrea "TheMiddle" Menin - App Sec Manager @ Sicuranext
Accedi
