Can I paste company source code into ChatGPT?

Q: Can I paste company source code into ChatGPT?

It depends, and a blanket ban is the wrong answer. Most code — boilerplate, framework glue, generic patterns — carries little risk. The risk concentrates in six things: client code under NDA, your core proprietary algorithm, security-relevant code such as authentication or payment logic, anything containing embedded secrets or internal hostnames, unpatched vulnerability details, and infrastructure-as-code with real account identifiers. Reduce the problem to a minimal example with renamed identifiers and fake data before pasting. The practical test: would you post this to a public Stack Overflow question under your real name with your employer in your bio?

It depends.

This is the page where a blanket ban would be wrong — and pretending otherwise is exactly why developers stop listening to security guidance. Most of the code you write isn't the secret. The framework is public, the pattern is on Stack Overflow, the algorithm is in a textbook.

The genuine risk is narrower and more specific than "don't paste code". Get that part right and you can use AI on code all day without a problem.

Where the risk actually is

Four distinct problems hide under the phrase "pasting code", and only some of them apply to any given paste.

Trade secret protection is conditional. Code qualifies as a trade secret only while you're taking reasonable steps to keep it secret. Disclosing it to a third-party vendor with no confidentiality agreement is a decent argument that you weren't. That's a legal position you can weaken by accident, in one paste, permanently.

Client code usually isn't yours to disclose. If you're an agency, a consultancy, or anyone working under an MSA, there's likely a clause naming who may process that code. An AI vendor is a subprocessor. Nobody added them to the list.

Code carries passengers. Hardcoded keys, internal hostnames, database names, private endpoints, a comment with someone's phone number. You pasted a function; you also pasted everything around it.

Architecture is intelligence. Internal service names, queue topology, which system talks to which — that's a map. Useful to you, and useful to anyone planning to attack you.

The narrow list: don't paste these

✕Client or customer code under NDA — not yours to disclose, regardless of tooling
✕The actual differentiator — the pricing engine, the matching algorithm, the model, the thing the company is
✕Security-relevant code — auth flows, crypto, session handling, payment logic, access control
✕Anything with secrets or internal names in it — see credentials; strip these before you do anything else
✕Unpatched vulnerability details — a live bug plus the code that causes it is an exploit, pre-disclosure
✕Infrastructure-as-code with real identifiers — account IDs, ARNs, VPC and subnet layout, security group rules

Usually fine

✓Open source you already publish — it's public. Pasting it changes nothing.
✓Boilerplate and framework glue — config, routing, plumbing that looks the same at ten thousand companies
✓Minimal reproductions with fake data — the twelve lines that actually fail, renamed
✓"How do I do X in Y" — questions with no company specifics in them at all

Do this instead

Reduce before you paste. The minimal reproducible example is a discipline developers already know from bug reports — it applies here unchanged, and it usually gets you a better answer anyway.

Don't

// 400 lines of
// PricingEngine.java

client = Client(
  host="pricing-prod.acme.internal",
  key="sk_live_9f2XkQ")

def acme_margin_ladder(tier):
  # our actual pricing logic
  ...

// 12 lines that reproduce it

client = Client(
  host="[HOST]",
  key="[KEY]")

def calculate_rate(tier):
  return rates[tier] * 1.2

// TypeError: unhashable type
// on line 3 when tier is None

Rename the domain concepts, fake the data, keep the structure that causes the bug. You've removed the confidential part and lost nothing the model needed.

The five-second test

Would you post this to a public Stack Overflow question, under your real name, with your employer in your bio?

If yes, paste it. If you hesitated, that hesitation is the answer — and it's the same instinct that made you reduce the snippet before posting it publicly. Nothing new to learn.

The exception

Approved enterprise coding assistants are a genuinely different position. Enterprise tiers of GitHub Copilot, Claude, ChatGPT and similar carry contractual no-training terms and, in some cases, run inside your own cloud tenancy. If your employer has assessed one and put it on the approved list, the calculus changes — that's what the assessment was for.

Two caveats that survive even then: client code under an NDA is still governed by the NDA, not by your tooling choice. And the secrets rule doesn't move — strip those regardless of which tool you're in.