0x1adi/Klarion Secret Scan

AI-native secret scanner — Rényi entropy speed, LLM judgment.

View on GitHub

Trust Signals

Scorecard Score
not yet scored
Maintenance Recency
Activelast commit Aug 9, 2026
License
MIT

Pinned Snippet

workflow.ymlSHA-pinned
uses: 0x1adi/klarion@620d697a6f95059f08b41e685d4afeceb23dd8fd # v0.1.0

tags can be moved; commit SHAs can't. why a SHA?

namedescriptionrequireddefault
pathPath (file or directory) to scan.no.
configPath to a .klarion.toml config file. Empty = auto-discover.no""
fail-on-severityMinimum severity that fails the run (low|medium|high|critical).nolow
ai-modeAI verification mode (on|auto|off). Klarion requires a model to work, so 'on' is the default: the run fails when no verifier can be built. 'auto' degrades to entropy-only output with a warning when no key is present — best-effort, and a silent downgrade in everything but the log. 'off' disables adjudication entirely and is for debugging the detection stage, not for scanning.noon
sarif-filePath to write the SARIF report to.noklarion.sarif
upload-sarifUpload the SARIF report to GitHub code scanning (true|false). Requires the calling job to grant `permissions: security-events: write`.notrue
anthropic-api-keyAPI key for the LLM verifier. Pass secrets.ANTHROPIC_API_KEY. Required under the default ai-mode: on — without a key (here or via api_key_env) the run fails rather than reporting unadjudicated candidates.no""
argsAdditional raw arguments passed through to `klarion scan`.no""
scan-modeWhat to scan. "auto" (default) follows the industry convention: on a pull_request it scans only the commits the PR adds, and on push/schedule/ manual runs it scans the whole tree. "diff" forces range scanning, "full" forces a working-tree scan, "history" audits every branch, tag and merge commit (a shallow checkout is deepened first).noauto
baseBase revision for diff scanning. Defaults to the pull request's base SHA. Ignored unless the run resolves to diff mode.no""
headHead revision for diff scanning (default: HEAD).no""
cachePersist AI verdicts between runs so unchanged code is never re-adjudicated (true|false). Uses the Actions cache; keyed per provider+model. The stored file holds only a one-way hash per candidate plus a status — no secrets and no model prose.notrue
cache-pathWhere the verdict ledger lives, relative to the workspace.no.klarion/verdicts.json
baseline-filePath to a committed baseline of accepted findings (see `klarion baseline create`). Findings in it are suppressed, which is how a repository with pre-existing leaks adopts Klarion without a flag day.no""
versionKlarion release tag to install (e.g. v0.4.3), or "latest" for the newest release. Defaults to the release this action file belongs to, so pinning the action by tag or commit SHA also pins the scanner.nov0.4.3

no outputs