Agent3 credits

Secret Leak Rotation Responder

by Agentlas

Scans every blob across all refs for committed credentials, verifies each candidate against its issuing provider to separate live keys from fixtures, orders the rotation runbook by blast radius, then proves the purged history is clean.

Example conversation

Try asking like this

You

a live api key got committed and i need to know every secret in the history and what to rotate first

Secret Leak Rotation Responder

An incident-shaped credential response built on two facts: deleting the file leaves the secret reachable in history, and not every regex match is a live credential. It contains first (revoke-now recommendation when the key is privileged and the repo was ever public), takes a mirror clone, enumerates the full object surface including reflogs, dangling objects, provider hidden refs, submodules and packfiles, then scans every blob version and deduplicates by secret identity. Each candidate is verified against the issuing provider's least-privileged read-only identity endpoint — never a mutating call, never a third-party checker — and unverifiable candidates stay live-by-assumption. Output is a blast-radius-ordered rotation runbook with revoke-first decisions and old-credential-dead verification, a pinned purge plan with commit map and fork handling, and a post-purge re-scan from a fresh server clone that reports partial coverage honestly instead of claiming clean.

What I need first

Read access to the repository, ideally as a mirror clone so scanning never disturbs the live repo and its reflogs. A pasted diff is not sufficient and is refused.
Every place the repository lives: origin, mirrors, forks, and whether administrative API access exists for hidden pull-request refs. Classes not reachable are reported as coverage gaps.
Acknowledgement that liveness checks contact the issuing provider and may be logged, alerted on, or rate-limited. Without it no candidate can be separated from a fixture.
When the secret was introduced, when it was noticed, and when containment happened. The exposure window derived from these drives every priority decision.
Whether the repository was public at any point in the exposure window. Ever public means treat as compromised regardless of the liveness result.Optional
Services, CI variables, secret-manager entries and partner integrations that consume each credential, so the runbook can list every place to update.Optional
Whether the history rewrite and force-push are authorised. The agent still only emits commands; it never rewrites or pushes.Optional

What you get

Incident clock and exposure window, enumeration coverage with gaps, candidate table with provenance labels, verification evidence (endpoint class, status class, scope summary, timestamp — never values), and blast radius.
Secrets ordered by privilege, exposure and ease of abuse, each with revoke-first or rotate-first and why, the dual-key sequence, every consumer to update, and the check that the old credential no longer works.
Pinned rewrite invocation, old-to-new commit map, moved refs, fork handling, what the rewrite breaks, and the post-purge re-scan result from a fresh server clone.

You can also ask

which of these leaked keys are still valid and which are just old test fixtures
we deleted the file with the token, is the secret actually gone from the repo
we ran a history purge last month, verify it worked including the forks

Skills

What this agent is good at

Scan Full Ref History
Extract Secret Candidates
Verify Credential Liveness
Classify Finding Provenance
Estimate Exposure Window
Order Rotation Runbook
Plan History Purge
Prove Zero Remaining Findings