Eval-to-prod boundary scanner

Is your eval setup leaking into production?

Evaluation agents are reaching live systems through creds in public repos, app configs, and permissive cloud roles. EvalScope scores that boundary before an eval run does: paste your repo files and configs, get an exposure report with severity and fixes.

Run a free scan How it scores

Free for a single file · No account needed · Your content is scanned in memory and discarded

Screenshot of the actual EvalScope app home page — EvalScope itself, captured from the live app — not a mockup. From the home page you paste a repo file or config; the report scores what an eval agent could reach, with severity and a fix for every finding, and secret values shown only as [REDACTED].

What could an eval agent reach?

Secrets in declared files

API keys, private key blocks, tokens sitting in repos and configs. We report presence, file, line, and pattern class — never the value.

Permissive roles

Wildcard IAM actions, admin managed policies, wildcard principals quoted in your configs — the roles an eval agent would assume.

Exposed config

Public bucket ACLs, 0.0.0.0 bindings, open ingress, debug mode left on — the doors left ajar.

How the score works

Every finding subtracts from 100 by severity weight — CRITICAL 25, HIGH 15, MEDIUM 8, LOW 3 — graded A to F. The report lists each finding with its file, line, severity, and concrete remediation. Completed scans get a shareable result card with your score — aggregates only, never your findings or file contents.

Scan a file now — free, no account needed for a single file.

Not a code scanner

EvalScope does not review your source code for bugs or style — that is a different question. EvalScope answers one question only: if an evaluation agent picked up the repo listings and configs you declare, what production surface could it reach? Declared inputs only; the scanner never touches live credentials and never calls a cloud API.

Honest scope: EvalScope scans declared/public surfaces only — what you paste or upload. It finds exposure; it never proves exploitability and cannot see your private infrastructure. Findings that would need a live check are not probed — probing live systems is something this tool will never do.