skip to content

~/how-it-works

Three passes from leaked to sealed.

Pattern work finds the candidates. Reasoning decides what they mean. Then you get the fix, in order.

$ leaksweep sweep

01

Sweep

Every credential-shaped string, found and masked in one pass.

  • >Input arrives as a paste, an archive, or a public repo URL and is held in memory only.
  • >Archives are filtered first: binaries, lockfiles, vendored and minified code are skipped, with per-file and total size caps against zip bombs.
  • >29 provider-specific and generic detectors run over each file. Generic hits must clear an entropy gate. References to env vars are dropped.
  • >Each hit is masked at source to first and last four characters, and given a keyed HMAC fingerprint. From this point on, nothing in the system holds the full value.
[ok] 214 files accepted, 1,902 skipped
[ok] 18,402 lines swept
[ok] 7 candidates, masked at source

$ leaksweep reason

02

Understand the real risk

The part that needs a language model.

  • >The reasoning layer receives masked values, five lines of redacted context, the file path, entropy, and the pattern engine's hints.
  • >It returns a verdict (live risk, likely real, false positive), a severity based on blast radius, and one sentence of reasoning for each.
  • >Pairing matters: an AWS key ID beside its secret is critical. A Stripe test key is low. A key in docs/examples with EXAMPLE in it is noise.
  • >Guardrails run after the model: structurally certain credentials cannot be dismissed by the model alone, and the risk score is computed in code.
[!!] 2 critical · 1 high
[--] 4 cut as noise, reasons attached
[ok] risk score 85 / 100

$ leaksweep fix

03

Rotate and prevent

A finding is only useful with its fix attached.

  • >Every live finding ships with ordered steps: rotate at the issuer, move to env in your stack's idiom, purge history, read the audit log.
  • >Commands come with copy buttons and placeholders in angle brackets. No part of a secret beyond the mask ever appears in one.
  • >After rotation, a re-sweep confirms the fix by fingerprint: sealed, still exposed, or new.
  • >The prevention layer closes the loop: a pre-commit hook and a CI gate, generated for the repo on Dev and Team.
[ok] rotation steps copied
[ok] re-sweep: 3 sealed, 0 remaining
[ok] secret-scan.yml generated

what never happens

The negative space matters more.

never shown

A full secret is never rendered. Not in the console, not in a report, not in an email.

never stored

No raw secret and no source code reaches storage. Masked value, fingerprint, location. That is the record.

never tested

Leaksweep does not try a credential against its service to see if it works. Validation is an attacker's move.

never trained on

Your code and findings are not used to train models, ours or a provider's.

More on the stance: security-first care. More on the model: how the AI works.

$ leaksweep init

Sweep your code before someone else does.

Three sweeps a month on the free tier. No card. Secrets masked from the first byte.