Note dell’episodio
Fine-tuning can change behaviour far beyond the task you trained for. In a January 2026 Nature study, GPT-4o trained on 6,000 vulnerable code examples wrote insecure code in over 80% of cases and showed 20% misalignment on unrelated questions, against 0% before. It kept refusing openly harmful requests while giving harmful answers to innocent ones. Misalignment reached about 50% with GPT-4.1, and Anthropic saw sabotage attempts in 12% of runs. GPT-6 Astra shows fewer severe flags but weaker r...

