Stop secret source code from leaking into ChatGPT, Cursor, and Claude Code.
Your developers will paste proprietary code into an AI assistant. Most of the time that code leaves your network and lands in someone else's model. Swiftward recognizes your own code and blocks it before the prompt ever leaves.
Who needs this
Not everyone. If your code is not a trade secret, a Terms of Service is probably enough for you. But if your edge lives in code, this is existential: high-frequency trading firms, exchanges and their matching engines, algorithmic-trading funds, chip designers with proprietary HDL. For them, a single pasted function can be the whole moat walking out the door.
How it works
This is a dedicated capability, not a regex. Swiftward parses your repositories, builds a fingerprint of your code, classifies it, and stores it. Every developer prompt is then matched against that fingerprint inline, before it leaves your environment. The match holds up through reformatting and light edits, not just an exact copy-paste, and it stays fast enough to run on every request. We will walk your team through how it works under NDA and measure false-positive and false-negative rates against your own repositories, so you verify it rather than take our word for it.
The threat is the honest mistake, not the spy
A developer with access to your code is not your adversary. They already have it; if they wanted to leak it, an AI tool is the last thing they would need. The real risk is the slip: someone opens a secret file in Cursor, or pastes a function from your matching engine into ChatGPT to debug it, without thinking about where it goes. Swiftward catches that and stops it before it leaves, so a moment of inattention does not become a leak. We do not try to defeat deliberate obfuscation or a heavy rewrite, because that is not the threat, and pretending otherwise would be dishonest. If code is refactored enough to look new, it is new; once it is committed back to the secret repository, the next index picks it up and it is covered again.
Kept current
Your code changes every day, so the fingerprints do too. Swiftward re-indexes what changed, on commit or on a schedule you set, so what is protected at runtime is what is in your repositories now, not a stale snapshot from onboarding.
Classify, then decide
You sort your repositories into sensitivity tiers and set what each may touch. The top tier never goes to any model. A middle tier can go to a model you host yourself but not to an outside one. The rest flow freely. On a match, Swiftward blocks the request or raises an alert, according to the tier and where the prompt was headed. There is no redaction here: a half-masked function is useless to the model and to you, so for source code the honest options are stop it or flag it, and you choose which per tier.
A rare and hard capability
There is no off-the-shelf library you can clone for this; it is hard to do well, and we built it as a real project of its own. It is one detector among many in Swiftward, feeding the same engine as every other rule, so it is versioned, testable, and audited like everything else. For the teams whose code is the business, it is the piece hardest to find anywhere else. Its sibling is personal data and secret leak prevention, which keeps PII and credentials out of a model the same way.