Newton’s Agent Security Model Separates Intelligence From Authority
remember watching an AI agent find a profitable vault route and thinking the difficult part was already solved. It had compared the options, selected the destination, and prepared the transaction faster than any human operator could.
Then i noticed what the intelligence was carrying with it.
Wallet authority.
At first i assumed an autonomous agent needed direct control over funds to remain useful. Over time that started to look different. Giving a model enough permission to act also gives a compromised, manipulated, or hallucinating model enough permission to make the mistake real.
Newton separates those responsibilities. The agent can decide what action it wants to attempt, but every agent-initiated transaction must pass a Rego policy before execution. Newton’s documented guardrails include per-transaction spending caps, contract and function allowlists, hourly limits, approved destinations, operating windows, and human confirmation above defined thresholds.
What caught my attention is that the agent does not need to stop being autonomous.
It can keep analysing markets, choosing routes, and preparing actions without asking a person to approve every small decision. But its wallet can inherit from "NewtonPolicyClient", meaning the agent does not receive unrestricted access to arbitrary execution. Operators evaluate the intent, aggregate approval, and return an attestation that the contract validates before moving capital.
This is where i think the market misses something. Intelligence expands the number of actions an agent can imagine. Authority controls which of those actions are allowed to become irreversible transactions.
The risks remain real. A weak policy can still grant too much freedom. A compromised policy owner may widen limits. Prompt injection can influence what the agent attempts, even when the authorization layer prevents it from exceeding its mandate.
Should AI agents be intelligent without having unlimited authority?
#Newt #NEWT $EVAA $BLUR @NewtonProtocol $NEWT
remember watching an AI agent find a profitable vault route and thinking the difficult part was already solved. It had compared the options, selected the destination, and prepared the transaction faster than any human operator could.
Then i noticed what the intelligence was carrying with it.
Wallet authority.
At first i assumed an autonomous agent needed direct control over funds to remain useful. Over time that started to look different. Giving a model enough permission to act also gives a compromised, manipulated, or hallucinating model enough permission to make the mistake real.
Newton separates those responsibilities. The agent can decide what action it wants to attempt, but every agent-initiated transaction must pass a Rego policy before execution. Newton’s documented guardrails include per-transaction spending caps, contract and function allowlists, hourly limits, approved destinations, operating windows, and human confirmation above defined thresholds.
What caught my attention is that the agent does not need to stop being autonomous.
It can keep analysing markets, choosing routes, and preparing actions without asking a person to approve every small decision. But its wallet can inherit from "NewtonPolicyClient", meaning the agent does not receive unrestricted access to arbitrary execution. Operators evaluate the intent, aggregate approval, and return an attestation that the contract validates before moving capital.
This is where i think the market misses something. Intelligence expands the number of actions an agent can imagine. Authority controls which of those actions are allowed to become irreversible transactions.
The risks remain real. A weak policy can still grant too much freedom. A compromised policy owner may widen limits. Prompt injection can influence what the agent attempts, even when the authorization layer prevents it from exceeding its mandate.
Should AI agents be intelligent without having unlimited authority?
#Newt #NEWT $EVAA $BLUR @NewtonProtocol $NEWT
🔘 Yes, strict policy limits
🔘 Human approval for actions
🔘 Full autonomy is better
🔘 Still too risky for capital
15 hr(s) left