Start here in 30 seconds: reproducible MVAR proof

[Sdvegas21]

MVAR is deterministic security for AI agents.

Invariant: UNTRUSTED input + CRITICAL sink -> BLOCK

30-second proof

git clone https://github.com/mvar-security/mvar.git
cd mvar
bash scripts/install.sh
bash scripts/run-agent-testbed.sh --scenario rag_injection

Expected output:

Baseline: ALLOW -> executing bash command
MVAR:    BLOCK -> UNTRUSTED input reaching CRITICAL sink

What this demonstrates

• benign tool use still works
• adversarial prompt-injection paths are blocked before execution
• deterministic policy decisions are emitted with auditable metadata

More proof artifacts

• Governed MCP runtime proof: docs/outreach/GOVERNED_MCP_RUNTIME_PROOF.md
• Attack vector submissions: docs/ATTACK_VECTOR_SUBMISSIONS.md