The short answer
Building with AI is as safe as the blast radius you design for. The blast radius test, from AI4NTP (AI for Non-Techy People) Session 008, asks one question: if this system's data or codebase ended up in a bad actor's hands, what would happen? Social security numbers demand buttoned-up security; a low-stakes internal dashboard does not need the same weight.
What is the blast radius test?
The blast radius test is a security-sizing question for any AI build: if this system's data or codebase ended up in a bad actor's hands, what would happen? Alec Saluga's observation on Session 008 was that many people use the cybersecurity argument as a reason not to build anything, when the honest move is to measure the actual consequence and secure to that.
The answer sets the budget. Housing customers' social security numbers means buttoned up without question, possibly with third-party security help. An internal dashboard whose leak would be a shrug does not need the same weight, and pretending it does is how projects die in committee.
What should you secure on day one?
Ian Kilpatrick's day-one checklist for any agent on a server has three moves. Raise the firewalls first. Install Fail2Ban second: it watches for failed logins and bans the offending IP for 24 hours after about 5 attempts, which dispels most of the automated attack traffic any public server attracts. Third, set the guardrails so the agent never puts API keys or passwords in the codebase or on GitHub.
None of the three costs money, and an agent can install all of them for you. The point of the checklist is that baseline hygiene is nearly free, which removes the last excuse for skipping it.
When is heavier security actually justified?
Consequence, not fear, justifies weight. A large blast radius, regulated data, customer financials, anything a bad actor could really hurt someone with, can justify bringing in third-party security help. AI4NTP has seen the far end of the spectrum in practice: a lawyer required a fully on-prem server because client data could not be exposed to any outside infrastructure at all. The same logic that lets a low-stakes dashboard ship fast is what demands the bunker when the radius is real.
Frequently asked questions
Is it safe to build with AI?
As safe as the blast radius you design for. Measure what happens if the system's data leaked, secure proportionally to that answer, and cover the free baseline everywhere: firewalls, Fail2Ban, and no credentials in the codebase or on GitHub.
What is Fail2Ban and do I need it?
Fail2Ban is free, open-source intrusion prevention that bans an IP after repeated failed logins, about 5 attempts for a 24-hour ban in Ian Kilpatrick's setup. Any agent living on a public server should have it; it is his first install after the firewalls.
When should I bring in security professionals?
When the blast radius is large: regulated data, social security numbers, anything whose leak gets really ugly. Below that line, the free baseline plus sensible guardrails covers most internal builds, which is exactly why the blast radius question is worth asking before the fear makes the decision.
Where this came from
58:00 Ianthe first thing I'll do is put it on a server and say, put up the... put up the firewalls, install Fail2Ban
58:00 Ianit'll ban them after 5 failed attempts or something like that for 24 hours
58:00 Ianjust make sure that it passes no API keys or, you know, nothing... nothing in the code base, nothing on GitHub that are passwords
59:00 Alecit's very important to look at what... what's the blast radius? What does the system have access to?
59:00 Alecif you have customers and you're housing their social security numbers, for instance, you gotta be buttoned up
60:00 Alecif the blast radius is large and could get really ugly, then you might need to bring in a third party to help out with security
60:00 Justinwe had a lawyer reach out to us that wanted to build a server on-prem, because of the sensitive PI
