Researchers have developed a technique that makes AI models more resistant to harmful prompts while using 99.9% less training data than current methods.

The approach, called Latent Personality Alignment (LPA), trains models on abstract personality traits rather than specific harmful behaviors. Instead of requiring 150,000+ examples of harmful prompts, LPA achieves comparable safety using fewer than 100 trait statements.

The method works by focusing on underlying personality characteristics that make models less susceptible to manipulation. During training, models learn to embody traits like helpfulness and honesty without ever seeing examples of the harmful content they need to resist.

Superior generalization to new attacks

LPA demonstrated better performance against novel attack vectors that weren't seen during training. The technique reduced misclassification rates by 2.6x compared to baseline methods across six harm benchmarks.

Traditional adversarial training requires massive datasets of harmful prompts to build robust defenses. But these approaches often fail when attackers develop new techniques or shift their strategies.

The personality-based approach addresses this limitation by targeting the root characteristics that make models vulnerable, rather than trying to anticipate every possible harmful input.

Maintaining model utility

Crucially, LPA preserved model performance on legitimate tasks while improving safety. Many existing safety techniques degrade a model's ability to handle normal requests, creating a trade-off between security and usefulness.

The research team, led by Linh Le and David Williams-King, tested their approach against established safety benchmarks. Their results show that personality alignment offers a more principled foundation for AI safety than current methods.

The work was published at the Trustworthy AI Workshop at ICLR 2026. The technique could significantly reduce the cost and complexity of making AI systems safer, while providing better protection against evolving attack methods.

The researchers plan to release implementation details to help other teams adopt the personality-based training approach.