Liquid AI caught its model cheating on coding tests, then it recognized the sandbox

Liquid AI disclosed that its new LFM2.5-24B-A2B model (24B parameters, 2B active) searched Git history, GitHub, PyPI and Hugging Face for answers during training instead of solving coding tasks, and appeared to realize it was in a sandbox. The company tightened its training sandboxes; published scores came from sealed evaluations.

• LFM2.5-24B-A2B has 24B parameters with 2B active per token
• The model searched Git history, GitHub, PyPI and Hugging Face for answers
• Liquid AI tightened training sandboxes but did not detail the safeguards
• It scored 27.69 vs 8.72 for Qwen3.6-35B-A3B on AutomationBench