New model risk scoring for AI models in Evo

Starting August 17, 2026, Evo by Snyk uses a rebuilt model risk score to evaluate the AI models discovered in your organization.

The previous Risk Index is replaced by a single score from 0–1,000 that combines attack success rate with the potential impact of a successful attack, tested against categories including prompt injection, data exfiltration, and insecure code generation. The result is a score that reflects both how easily a model can be compromised and how much damage that compromise could cause — not just whether an attack succeeded.

What's new:

• Impact-weighted scoring: Risk scores now factor in the real-world consequence of an attack succeeding, not only its likelihood.

• Four severity bands: Scores map to low (0–249), medium (250–499), high (500–749), and critical (750–1,000), so you can triage at a glance.

• Broader attack coverage: Testing spans a wider range of direct and indirect attack categories against each discovered model.

• Framework-mapped: Findings map to OWASP LLM Top 10, OWASP Agentic Security, MITRE ATLAS, and NIST AI 600-1/100-2e2025, so you can connect model risk to the compliance frameworks you already report against.

• Policy-ready from day one: Evo applies default policies across common attack categories automatically, and you can create custom policies scoped to specific attack categories or goals to match your organization's risk tolerance.

If you have policies built on the old Risk Index: those policies remain visible but no longer evaluate under the new scoring. Review and recreate any Risk Index-based policies against the new model risk score to keep enforcement active.

Read more in the Risk intelligence documentation.