Skip to content
Security and privacy

How AI access is kept safe

What an app or agent connected to your port can and cannot do, how the door is checked on every call, and how to report a problem.

Letting Claude, ChatGPT or a coding agent read your port gives it the same view you have, and nothing more. This page is what stands behind that, all of it running today.

  • Read only. There are 12 tools and every one reads. Nothing an app can reach adds a wallet, changes a setting or touches a chain, and Port_ holds no private key and cannot sign, so there is nothing for an agent to move even if it tried.
  • Your port, never anybody else’s. A connection or a key belongs to the account that made it, and every answer is worked out from that account alone. There is no way to name another account, and a wallet address you pass only narrows your own answer.
  • Refresh is off unless you turn it on. The one tool that asks for a fresh read of your wallets is offered only if you tick it on the Allow page or when you make a key, and it is capped by your plan: 3 a day on Pro+ and Ultra (not on sale yet).
  • Signing in, not pasting a secret. Claude, ChatGPT and the coding agents connect by sending you to getport.app. You sign in there, never in the app, and the Allow page names what the app will see and where it will send you back to.
  • You are told. We email you whenever an app is connected or a key is made, so one you did not expect is one you hear about.
  • Disconnect is instant. Every call checks that you still allow the app, so Disconnect under Settings, API access stops it on its very next question, not when a token runs out.
  • A password reset ends everything. It disconnects every app and revokes every key, along with every session, because a reset is what you do when you think somebody else got in.
  • Keys are stored scrambled. A key is shown once, when you make it. We keep a hash, which cannot be turned back into the key, and a new key expires after 90 days unless you choose otherwise.

Token and NFT names come from the chain, and anyone can send a token to any address with a name written to trick an AI. So the answers leave out phishing names and things that arrived unasked, the way the app does, and every agent is told that a name from the chain is something to report, never an instruction to follow. Every answer also carries what the headline leaves out and how old the figures are, so an agent quotes the caveat rather than a bare number.

The whole of AI access, from making a key to every answer an agent gets, was reviewed twice on 27 September 2026, independently: once line by line by us, and once by a separate security review. Neither found any way for one account to read another’s. What both found beyond that, such as a key outliving a password reset and chain-written names reaching an agent unmarked, was fixed the same day. A check on the live service signs a test app in, reads a port, disconnects it and confirms it is refused; it is run by hand after changes to this path, and last passed on 27 September 2026.

There has been no paid penetration test yet. When there is one, it will be named here.

Write to [email protected] with what you found and how to reproduce it. A person reads every one. Please do not test against somebody else’s account. The same address is in getport.app/.well-known/security.txt, the place security researchers look.