AI safety evolved: secure-by-design, safe-by-measurement · Blog · Sublime Security

This Sublime blog post discusses how adversaries can target AI models (not just users) via prompt injection and outlines a maturity checklist and four AI-safety pillars—adversarial input, calibrated output, tiered autonomy, and symmetric red-teaming—to safely operate autonomous email-security agents. It details baseline, higher-autonomy, and advanced operational controls, current practices (guardrails, trust boundaries, calibration, red-teaming), measurement approaches (calibration curves, Brier scoring, canary sets), and roadmap items such as publishing taxonomies, calibration reports, and circuit breakers.