What happens to your code.

What we take in

The input to an audit is a source tree at one fixed commit. There is no field for a URL, host or other live target, and nothing is scanned or tested over a network. We never contact your systems.

Isolation

• A separate container for every job. Each reviewer and each test run gets its own container, running as a non-root user with a read-only filesystem and all extra privileges dropped. The container host’s Docker socket is never exposed to it.
• Separate networks. Each job runs on its own internal networks, and one job’s containers cannot reach another’s.
• A copy, not the original. Containers receive a copy of your code with the git history stripped.
• Your repository cannot configure the reviewers. Their settings and credentials are kept outside the code under review.

Network access

There are three kinds of network access, and each is separate.

• Dependency install: official package registries only, no AI, no API key. Dependencies are installed before any reviewer sees your code, through a proxy that allows only the official registries for your ecosystem (such as npm, PyPI, crates.io, the Go proxy, Packagist and RubyGems) over HTTPS. Where a package manager supports turning off install scripts, they are turned off. Some package managers (for example pip and bundler) have no such switch; for those, install scripts can run, in this registries-only step that has no AI and no key. The container is then discarded, and the tree is rebuilt from your original checkout.
• Review: the AI gateway only. Reviewers can reach nothing except a gateway that forwards model requests to Anthropic. The gateway refuses web search, web fetch and similar tools, and refuses requests that point the model at a URL, so the model cannot be made to fetch from or send your code to a third party.
• Tests: no network at all. Every regression-test run has no network access.

The API key is held only by the gateway. Reviewers cannot read it; each receives its own token, revoked when the job closes.

Who sees your code

Client code is sent only to Anthropic’s commercial API, which does not use it for training. We train no models and have no training pipeline. Your code is not sent anywhere else.

Retention

Your code is deleted when the audit ends, and a scheduled sweep removes anything older than 24 hours.

Deletion covers the checkout, the installed dependencies, working copies and the reviewers’ transcripts. What remains is a small job record (identifiers and status) with agent-written text scrubbed out, and the report delivered to you. After deletion the system checks that nothing is left and alerts the operator if anything is.

Secrets found in your code

Reviewers are instructed never to copy a secret into their output. A finding refers to a secret by its location, its type, a mask showing only the first 4 and last 4 characters, and the first 12 characters of its SHA-256 fingerprint. Automated checks reject a draft that contains a secret value, and the report generator masks secret-like strings as a further backstop.

What this page does not cover

Terms, payment and the contractual side of data protection are being finalized; see Terms and Privacy.

Questions about your code?

Ask in your quote request and we will answer before you share anything.