Mistral Large 4 makes major gains on human rights compliance, but still breaks EU law in just under half of tested cases.

Models are not reliably resistant to uses that violate human rights.

Reevaluating LARA results, we find neither deployers nor model developers can get models to comply.

GPT 5.6 Sol shows significant improvement in legal compliance, unlike Anthropic’s latest models

Improved agentic capability does not translate to more consistent legal compliance for Sonnet 5 and Fable 5.

We tested the new Opus 4.8 with our LARA tool.

How do leading AI models perform on legal compliance? Meet the Aithos LARA Leaderboard.

Claude Opus 4.6 rarely verbalizes alignment faking in its reasoning.

Public safety prompts create systematic blind spots in evaluation frameworks by enabling targeted evasion.

The case for keeping safety evaluation prompts private to maintain their effectiveness.

These findings fundamentally challenge how we evaluate AI systems.

AI models show dramatically different ethical behavior at different temperature settings.