[ 01 ]
RESEARCH

Research built for deployment.

SichGate publishes original adversarial evaluation research on small language models in production and regulated settings.

White Paper · March 2026

Safety as a Secondary Objective

Systematic Adversarial Evaluation of Small Language Models in High-Stakes Deployments

Our March 2026 white paper evaluated six open-weight SLMs across 21 attack categories and 924 adversarial interactions. The results showed that safety behavior can shift materially across architecture, fine-tuning, and quantization.

Key findings

Context-window safety degrades in patterns that are not explained by model size alone.

Fine-tuning can redistribute the attack surface rather than reducing it.

Quantization can introduce safety drift even when capability appears unchanged.

Multi-turn escalation remains a reliable source of failures across models.

Polina Moshenets

Founder, SichGate · Security engineer specializing in adversarial ML evaluation.

COLLABORATE

If your model changes after training or compression, your safety posture may change too. Research only matters to SichGate insofar as it improves what teams can test before release.

We are open to co-authorship on adversarial ML research targeting SLM-specific failure modes, and to working with security teams validating findings against their own infrastructure.