Formal safety framework for AI agents. Pluggable LLM reasoning constrained by mathematically proven budget, invariant, and termination guarantee. 7 theorems enforced by construction, not by prompting. Includes Bayesian belief tracking, causal dependency graphs, sandboxed attestors, environment reconciliation, and a 155-test adversarial suite.
python cybersecurity bayesian-inference formal-verification autonomous-agents ai-safety agentic-ai agentic-ai-security agent-safety-frameworks
-
Updated
Mar 4, 2026 - Python