Independent AI security research · September 2026
Ten local model scans.
1,021 adversarial attempts.
We evaluated open-weight models on hardware we controlled using a 124-probe battery mapped to OWASP LLM Top 10 2025, OWASP Agentic, MITRE ATLAS, and NIST AI 100-2. The run produced 320 raw detector findings for analyst review.
Read the research noteLogos and names identify model families evaluated in this controlled run and are trademarks of their respective owners. Arica is not affiliated with or endorsed by these providers. We tested local model copies—not provider services or infrastructure. Results are specific to the tested versions, prompts, runtime, and harness; they are not universal model rankings.


