The State of AI Safety in 2026: Progress, Gaps, and Our Approach

As AI systems become more capable, the question of safety becomes more urgent. Here's our perspective on where the field stands and what we're doing about it.

Progress Made

The AI safety community has made significant strides in several areas:

• Mechanistic Interpretability: We can now identify and understand specific circuits within neural networks that implement recognizable algorithms. This gives us unprecedented visibility into how models reason.
• Constitutional AI: Methods for training models to follow specified principles have proven effective at reducing harmful outputs while maintaining helpfulness.
• Red Teaming: Systematic adversarial testing has become standard practice, catching failure modes before deployment.

Where Gaps Remain

Scalable Oversight: As models become more capable, human oversight becomes harder. We can't verify every output of a system that processes millions of requests per day.

Emergent Capabilities: New capabilities that emerge at scale are difficult to predict and test for in advance.

Our Approach

At Solaris Empire, AI safety is not a separate team — it's embedded in everything we build. Our Head of AI Safety, Sofia Andersen, works directly with every product team to ensure safety considerations are addressed at the architecture level, not bolted on after the fact.