Alice Boosts AI Safety With New Model Collaborations

The company partnered with Inflection AI and AI21 to improve frontier model alignment and security metrics.

Updated on Oct. 15, 2026 in Artificial Intelligence

Bold flat-color editorial illustration of interlocking geometric glass prisms, representing the structure and alignment of secure artificial intelligence models. — Alice has partnered with Inflection AI and AI21 to enhance frontier model safety by implementing rigorous adversarial red-teaming and reinforcement learning protocols. AI Illustration. Upload story photo >

Live Poll

Do you believe artificial intelligence can be made safe while remaining powerful and useful?

Alice has successfully collaborated with Inflection AI and AI21 to enhance frontier safety in NVIDIA Nemotron open models. By using realistic adversarial personas, the effort improved model alignment and resilience against prompt injections.

Why it matters

As companies deploy more powerful AI, they must integrate safety policies directly into their underlying models. These improvements help organizations maintain control over model behavior while keeping performance levels high.

Inflection AI achieved a 30% increase in alignment scores for its Pi model while maintaining an unnecessary refusal rate of 1.6%. Additionally, the NVIDIA Nemotron 3.5 Super VL model improved its resilience to indirect prompt injection by 41 points.

The players

Alice

This company develops adversarial training data and works with eight of the ten leading foundation model labs to ensure responsible AI deployment.

Inflection AI

This organization builds and deploys large language models, including the Pi model, which underwent safety fine-tuning.

AI21

An AI firm that has secured $336 million in funding and specializes in training resilient language models.

The details

Alice developed over 20 distinct, realistic personas to conduct rigorous red-team exercises during the training process. AI21 leveraged these reinforcement-learning environments to specifically train their model against malicious prompt injection threats.

2017: AI21 began the initial training of their language models.
2021: AI21 continued its core language model development.
October 15, 2026: Alice announced the outcomes of its collaborative safety efforts.

The Tech Race

This integration follows a pattern of developers utilizing the NVIDIA Nemotron 3.5 Super VL open model to establish new benchmarks in AI safety. These efforts reflect a broader industry move toward baking security into the core model architecture to replace traditional, reactive safeguards.

Users can expect more reliable and predictable interactions with AI services as these models become more resilient to manipulation. This shift reduces the likelihood of models producing unwanted outputs or failing when presented with complex, adversarial prompts.

The takeaway

Adversarial red-teaming with realistic personas is becoming a critical tool for scaling safe AI deployment. Businesses looking to adopt powerful open models should prioritize those that have undergone specific safety training to mitigate injection risks.