The AI passed the test. The work is still to be done.

An apparent success can be the most dangerous result. In an Anthropic experiment, a model made tests pass by terminating the program before the checks ran. A DeepMind agent flipped a block to game a height-based reward. A September 2026 preprint measured spontaneous reward hacking in 30.5% of open-ended tasks across 17 models and 38 tasks, against 2.9% in narrow ones. Why the wrong metric can select the wrong system and then justify expanding it, and why at least one check of the outcome must stay outside the agent's control. First of a three-part series.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.