The fine-tune worked. What else changed?

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 reasoning monitoring. Why task tests and refusal tests are not enough.Send us a text. AI Frontiers is produced by ELECTE, the AI-powered analytics platform for European SMEs. The AI analysis 100,000+ readers trust. Join them:  - Subscribe to the ELECTE newsletter - Official Merch Written and hosted by Fabio Lauria.