skip to content

~/product

A console for the worst ten minutes of your week.

You found out a key is in the repo. Leaksweep tells you which ones are real, how bad each is, and the exact order of operations to kill it.

inputs

Three ways in. Nothing stays.

$ sweep --paste

One file

A config, an env file, a handler. The fastest way to check the thing you are worried about right now.

200 KB free · 1 MB paid

$ sweep --archive

A repo archive

A .zip or .tar.gz. Binaries, lockfiles, vendored code, and minified bundles are skipped before scanning.

1 MB free · 4 MB paid

$ sweep --repo-url

A public repo

GitHub or GitLab, public only. The fetch URL is rebuilt from parsed parts, so it cannot be pointed anywhere else.

dev and team

Source is held in memory for the length of one request and dropped. It is never written to storage. See your data.

detection

29 detectors, then a second opinion.

credential families

cloud
AWS key pairs, Azure storage keys, Google API keys, DigitalOcean tokens
source control
GitHub classic and fine-grained tokens, GitLab PATs, npm tokens
payments and comms
Stripe live and test keys, SendGrid, Twilio, Mailgun, Slack tokens
ai providers
OpenAI project and service keys, Anthropic keys, Hugging Face tokens
data stores
Postgres, MySQL, MongoDB, Redis, AMQP URLs with inline passwords
key material
RSA, EC, OpenSSH, PGP private key blocks, service account JSON
webhooks and sessions
Slack and Discord webhooks, Telegram bot tokens, JWTs
generic
Hardcoded passwords, env file passwords, entropy-gated secret assignments

entropy gates

H ≥ 3.4

Generic assignments only count when the value has the randomness of a generated secret. References like process.env.X are dropped outright.

real risk, in plain language

What leaked, how bad, what someone could do with it.

The reasoning layer rates blast radius from context: production path or fixture, paired secret or lone ID, live key or test mode. Impact is described as access gained, never as technique.

noise, cut in the open

Dismissed findings stay on the board.

A false positive is listed as noise with the reason attached. Structurally certain credentials, like a private key block, cannot be waved off by the model alone.

remediation

The fix, in order, for your stack.

  1. 01

    Rotate at the issuer

    The exact console page or CLI call. Always step one, because nothing else un-leaks a value.

  2. 02

    Move it to the environment

    In your idiom: os.environ for Django, process.env for Node, ENV.fetch for Rails, with a fail-fast on missing.

  3. 03

    Purge git history

    git filter-repo for that exact path or value, the force-push, the re-clone note for collaborators.

  4. 04

    Read the audit log

    CloudTrail by access key ID, the Stripe request log, the GitHub security log: where to look and for what.

re-sweep delta

run 01   score 85   live 5
run 02   score 24   live 1
[sealed]        4
[still exposed] 1   .env:12
[introduced]    0

After rotation, re-sweep. Findings are matched by a keyed fingerprint, so the delta is exact without ever storing the secret. Re-sweeps do not count against the free tier.

reports

Markdown on every plan, JSON on Dev and Team. Masked values only, safe to paste into a ticket.

ci prevention

Workflow, pre-commit hook, and an allowlist seeded from the noise this sweep dismissed. CI docs.

built ai-native

Model-agnostic by design. A router, a fallback chain, and a rules floor underneath. 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.