Autonomous Risk Containment

Center Guard Policy Studio & Monotonic Attenuation Engine

Mandant solves the autonomous agent containment crisis by enforcing permissions directly in EVM smart contract bytecode. Child agents can only narrow permissions, never expand beyond their ancestor mandate.

Security PostureMAXIMAL

EVM GuardrailBYTECODE

Interactive Policy Configuration

Drafting Policy

Max Spend Per Job Ceiling$250 USDC

$50 (Micro-Job)$500 (Batch Inference)$1,000 (Max Capacity)

Permitted Execution Tools

AI Compute & GPU Inference

ERC-8183 escrow settlement for AI nodes

Web Crawling & Vector Pipelines

Deterministic dataset procurement

DEX Swaps (Uniswap / PancakeSwap)

Default: Disabled for autonomous workers

Raw ERC-20 transfer()

Hard-coded bytecode revert barrier

HARD-BLOCKED

Onchain policyHash Commitment:

0x7a3e0661098aa25e09592fd34a90ff5a8921d27162bf011fdb75d7c7b147a6f5

Committed into the MandateNode struct slot in MandateTree.sol on Arbitrum Sepolia.

Interactive EVM Bytecode Simulator

TEST CALLS

Test how the EVM executes when an autonomous worker calls different actions against this policy:

Click any scenario above to execute bytecode simulation...

Monotonic Attenuation Verification

In Mandant's recursive hierarchy, when Orchestrator (Depth 1) spawns Worker Beta (Depth 2), the smart contract enforces:

Allowlist(Child) ⊆ Allowlist(Parent) ∧ Budget(Child) ≤ Idle(Parent)

If an operator or agent attempts to pass an address outside the parent allowlist, the EVM transaction reverts instantly with ChildAllowlistNotSubset().