Sudont is a drop-in RPC proxy that executes every transaction inside a local dual-engine sandbox — revm for EVM, LiteSVM for SVM — before it reaches the chain. Simulation runs inline with your RPC call, not a round-trip to a SaaS cluster, so it completes inside the agent's decision window with no perceptible overhead. Prompt quality keeps improving; adversarial code and market physics do not. Sudont defends the surface that isn't going to get better on its own.
The last generation of agent safety treated a signing problem as a comprehension problem — better prompts, better reviewers, better intent verification. That class of risk decays: models get sharper, guardrails get stricter, prompt injection shrinks. The remaining risk does not. MEV alone has been "solved" a dozen times and still drains billions from agents that understood their intent perfectly and signed anyway. Autonomous swarms execute at machine speed across EVM and SVM, and the chain cannot tell a mathematically optimal trade from a hallucinated one. Every transaction is a potential Mythos-class attack — MEV sandwiches, smart contract trapdoors, hostile-state recursion — that probabilistic validators never catch. Intent cannot be verified. Only physics can.
Sudont is a drop-in RPC proxy that simulates every transaction against real execution physics before it touches the chain. No custom SDKs. No LLMs in the hot path. Simulation is deterministic, local, and finishes inside the agent's decision window.
Other transaction-safety tools are dashboards for humans to review risky transactions before broadcast. Sudont is an execution surface for machines to reject them. Inline, not alongside. Physics, not heuristics. Deterministic, not probabilistic. Local, not remote.
Read the Docs →See the dual-engine RPC proxy block a Mythos-class attack in real time. Sudont is onboarding institutional teams running autonomous agents on EVM and SVM. Capacity is intentionally constrained.
Or email sudo@sudont.xyz