Skip to content
Watchdog
Sign inSurvey a repo free

Guide

Answers for your security reviewer

Watchdog asks for read access only, reads your code in memory on its own servers in Denmark, and deletes its copy when the survey ends. What you keep is a signed report, and anyone you share it with can check that it is genuine and unchanged, without an account with Watchdog. This guide answers the questions a security review asks, in the order it usually asks them, and links to the contract and the checks behind the answers. All guides

Who this is for: security reviewers, and the technical buyers and engineering leaders who need their approval before connecting a repository. It covers what Watchdog can reach, where your code is analysed and what happens to it afterwards, which AI is involved, who can see the results, what runs during a survey, and how to check the signed report you receive.

Between surveys, there is no copy of your code at Watchdog to protect. Watchdog holds the reports, which quote only the few lines each finding points to, and a read-only connection to your code host, which you can revoke there at any time.


1. What Watchdog receives

How Watchdog reaches a repository depends on where it is hosted, and every route gives it read access only. On GitHub, you install the Watchdog GitHub App and choose which repositories it can see, and it asks only to read their code and basic details (Contents and Metadata, read-only). On GitLab and Bitbucket, you authorise Watchdog from your account, and Watchdog keeps the read-only access token this gives it, encrypted (on GitLab, the scopes are read_api and read_repository). On Azure DevOps, you paste a personal access token for one organisation that needs only permission to read code (Code (read)), and Watchdog checks that the token can read that organisation's code before it saves it. On all four, you can add a public repository by its address, with no token at all.

None of these connections can change your code, and the update messages your code host sends Watchdog (webhooks) carry no code. Signing in is separate from all of it. You sign in with your GitHub or GitLab account through Watchdog's own sign-in service, which runs on Watchdog's servers (Keycloak), and Watchdog receives only your user name, display name and email address. Signing in gives Watchdog no access to any repository. The access that reads code belongs to the repository connection, so you can sign in with GitHub and survey a repository on Azure DevOps.

On GitHub, Watchdog can also keep a single issue in your repository up to date with the open findings (the audit-marker issue). It is off until you turn it on, and it needs the App to have permission to write issues on that repository (issues:write). It creates, updates and closes that one issue and nothing else, and the rest of Watchdog works the same without it.

Two other connections are yours to add, and both are off until you set them up. On GitHub, you can let Watchdog read your code scanning results (security_events, read-only), so that each survey shows where Watchdog's findings and your scanner's agree. Watchdog keeps only those counts, never your scanner's findings. And Watchdog can create tickets for the work in your task tracker, such as Jira, Linear, Azure Boards or GitHub Issues, using a separate token that you create for that alone.

For code that is on none of those four hosts, you can upload it as a ZIP file for each survey.

Through every one of these connections, Watchdog only reads your code. It makes no commits, branches, pushes, pull requests or file edits.


2. No part of Watchdog can change your code

Each survey runs in one direction: Watchdog copies the code, analyses it, signs the result and delivers the report to you. It needs nothing installed in your repository or your build pipeline, and it runs alongside the linters, security scanners and build checks you already use.

Watchdog's tools for coding agents follow the same rule (an MCP endpoint, the standard way AI agents connect to outside tools). Through them, an AI coding agent or a person can ask what the last survey found and what to fix first, claim a piece of work so that two agents don't take the same one, mark it fixed, or argue that a finding is wrong. No tool writes to your repository, and no tool lets an agent raise its own score. A finding an agent marks as fixed counts as fixed only when the next survey no longer finds it. A finding an agent says is wrong is sent, with the agent's reason, to a person at Watchdog, who corrects the check if the finding was wrong. Each agent key works on one repository and can do only what you allowed it, and an agent can ask for a new survey only if the repository's owner has turned that on. Your agent does the fixing, and only the next survey decides whether it worked.

The compliance report works the same way. On a product's compliance report, a requirement that Watchdog caught failing stays open until a survey shows it fixed, and nobody can tick it off by hand. A person at your company declares the rest and signs the report, so your auditor can see which answers Watchdog measured and which a person declared.


3. Where the code is processed, and how long each thing is kept

Every run works from a copy that exists for that scan. Two questions decide whether that sentence means anything: where it lives, and when it goes.

Where. The scan checkout and the engine's scratch space are RAM-backed tmpfs mounts, asserted on every deploy: a host whose mounts, cleanup timer or encrypted volume have drifted fails the release rather than serving from a state the data-processing agreement does not describe. Your source tree is never written to disk. Analysis runs in a container with Linux capabilities dropped and no new privileges, on a deny-by-default egress network. Cloning runs in the scan worker outside that container, writing into the RAM-backed path.

How long, by data class:

  • Cloned repository source. RAM only. Deleted when the scan finishes, whether it completed, failed or was cancelled. A recurring sweep removes anything a crashed run leaves behind, and nothing survives a restart of the host.
  • Uploaded ZIP source. RAM only. Not deleted at the end of the scan, because unlike a repository it cannot be fetched again, and dropping it would mean asking you to upload a second time. Removed after seven days without use, and gone on restart either way.
  • Per-scan working data. RAM only. Deleted when the scan finishes, including on failure.
  • Reports, findings, scores, trend points, run metadata. Encrypted volume (LUKS2, AES-256-XTS). Retained while the repository is connected.
  • Account data. Database. You can close the account yourself. There is a 14-day reversible period, after which it is deleted. Invoices are kept where bookkeeping law requires.

Read the fourth entry carefully, because it is where "we do not keep your code" is most easily over-read. Reports quote short code excerpts: the file path, the line number and the fragment a finding cites. That is derived material about your source and it persists. What does not persist is the tree.

The clone credential is used only for the clone or fetch and is stripped from the checkout's stored remote, so it does not sit on any medium afterwards. On GitHub it is minted fresh and short-lived for each scan from the App installation. On the other three the credential is the access token or PAT you authorised, held encrypted and used the same way. Standing access is the connection itself, and you can revoke it at any time, either at the code host or by disconnecting the account here.

If your posture requires that source never leave your own boundary, a self-hosted deployment does the clone and the analysis on your infrastructure and sends only results outward. Its ingestion path is token-authenticated per repository.


4. The language model runs on our own hardware

Some findings carry an advisory read from a language model, used for prioritisation and explanation.

In the managed service, that model is self-hosted on hardware Canine Development owns and operates. No source code is sent to OpenAI, Anthropic, Google or any other AI provider, and your code is not used to train any model, ours or anyone else's. This is structural rather than a setting: the analysis container's only route out is an allowlisting proxy that permits package registries and scanner databases, plus a host-side endpoint whose upstream is the local model. No AI-provider domain is reachable from the environment your code is analysed in.

A customer running Watchdog on their own infrastructure configures their own inference endpoint, and the choice is theirs.

Two consequences. On residency: all processing and storage happens on servers owned and operated by Canine Development in Denmark, so your data stays in the EU. Note that your existing code host remains your own relationship and is not relocated by this - it is a supplier you chose, governed by your agreement with it. On the score: because the model is advisory, its output is marked as advisory in the artifact and can change what a report says but never the computed number. The measurement is reproducible independently of any model.


5. Who can see what is kept

On GitHub, report access mirrors your access at the code host, re-synchronised daily: access lost there is access lost here within a day. On GitLab, Bitbucket and Azure DevOps it does not, and a reviewer should know why rather than assume otherwise: the read-scoped credential those hosts give us can read code but cannot list who else has access, so there is nothing to mirror. On those hosts, access here is whatever your Watchdog team membership and per-person grants say. Removing somebody at the code host does not remove them here; removing them from the team does, immediately. Public reports exist only for public repositories that are published, and they pass through a sanitiser that strips account names and security-sensitive specifics. Production access inside Canine Development is restricted to named operations personnel and is logged. There are no contractors and no third parties in that path.

The whole sub-processor list is one entry: a payment processor, which activates only when billing is taken into use. If you do not use billing there are no sub-processors at all.


6. The deliverable is a signed package you hold

Every scan produces a CAI delivery package: a content-addressed manifest with an Ed25519 signature, pinned to the exact commit and the rubric version, with the supporting evidence embedded. The reports travel inside it - Executive, Engineering, Security and Audit views as HTML and PDF, alongside CSV, SARIF and CycloneDX SBOM outputs.

Two checks confirm one of these, and they are separate:

  1. The signature. Ed25519-verify the package against the issuer's public key. A valid signature proves the package was produced by the issuer rather than assembled by whoever shared it.
  2. The content hash. Recompute the hash over the manifest and compare. One edited byte breaks it.

Both are documented at codeassuranceindex.info/verify, and real signed packages are browsable at watchdog.canine.dev/publicreports. Neither check requires an account with us. Separately, the score itself is reproducible: run the open reference scorer over the published evidence, pinned to the same rubric version, and you should get the same number.

Reproducibility is a property of the record, not a promise about it. Every report carries the audit trail behind its deep-scan dimensions: each external tool, the exact command, the number of findings it returned, and a link to the retained raw output. Every command is repo-relative, so checking out the same commit and running the line shown reproduces the finding.

Two rows, verbatim, from the audit trail of the published survey of psf/requests. Each names the dimension, the tool, its captured version, the exact command, how many findings it returned, and a link to the retained raw output:

  • D28 Secrets (history) - gitleaks - gitleaks detect --no-banner --report-format json --report-path stdout --exit-code 0 --source .
  • D29 Static Analysis (SAST) - semgrep - semgrep --config /opt/semgrep-rules/security-audit.yml --config /opt/semgrep-rules/owasp-top-ten.yml --json --quiet

Every command is repo-relative rather than an absolute scratch path, so checking out the same commit and running the line shown reproduces the finding. A dimension with nothing to analyse records that, and why, in the same table. The full report carries the whole trail, and per-invocation exit codes and wall-clock durations sit beside it in the run's sidecar.

This is what makes the read-only posture worth something. You are not renting continued access to a system that holds your code; you are holding a document that stands up on its own.


7. What Watchdog does not do

  • Read-only on your repository, which is the claim rather than "we never run anything". It reads source, history, configuration and manifests, and where the toolchain allows it builds the project and runs its test suite inside the isolated analyzer: that is how build integrity and test coverage are measured at all. It never writes to your repository, meaning no commits, no branches, no pushes and no edits. It does not run your application, probe a live endpoint, perform dynamic scanning or conduct a penetration test.
  • Evidence, not certification. It produces signed, reproducible evidence about the automatable part of code quality and security posture. It does not make you compliant and does not certify you against GDPR, SOC 2 or ISO 27001. The division of labour is that the engine measures what can be measured, you declare the rest, and a named human signs. Canine Development holds no SOC 2 attestation today.
  • The codebase is the unit of measurement. There is no individual-developer score and no velocity or DORA-style ranking. Where dimensions read git history they identify where knowledge is concentrated, and no per-person score is computed or stored.

Coverage: the production catalogue currently holds 17 languages, with JavaScript analysed structurally rather than deeply. The catalogue is a live production record and may change, so read the current one at watchdog.canine.dev/api/public/language-support rather than this sentence. The current rubric, rubric-2026.10.1, holds 165 catalogue entries across ten lenses, five always on and five that light up when the codebase warrants them; the live catalogue is at codeassuranceindex.info/dimensions and the standard at codeassuranceindex.info/spec.


8. The short answers

Where does our source code go, and how long is it kept? Into RAM on servers in Denmark, for the length of one scan. A cloned repository is deleted when the scan finishes; an uploaded ZIP is kept until seven days without use, because it cannot be re-fetched. Neither survives a restart. What persists is the report package and the short code excerpts its findings cite, on an encrypted volume, for as long as the repository is connected.

Do you send our code to a third-party AI? No. In the managed service the model is self-hosted, and the analysis environment's egress allowlist does not include any AI provider. A customer running their own deployment configures their own endpoint.

Do you train on our code? No - neither our own models nor anyone else's.

Can Watchdog change our code or our repo? Not your code. There are no commits, branches, pushes, pull requests or file edits, and no MCP tool writes to a repository. One optional feature, off until you enable it, maintains a single GitHub issue as an audit marker and needs issues:write to do it.

Is this a penetration test? No. It is source-side analysis. Dynamic and outside-in testing are a separate discipline.

Does this make us compliant or certified? No. It produces evidence. You declare the rest, and a human signs.

It all runs on hardware you own, in one place. What happens if you lose it? The application is stateless and comes back in minutes. The databases are dumped twice a night by two independent jobs, one on the machine and one shipped off it with the configuration and the wrapped credential ring, and each dump is verified as it is written, so the exposure is up to 24 hours of data with no point-in-time recovery between dumps. The gap that remains is the site: the second machine is in the same building, so an off-site copy is still outstanding, and it is written down as a gap rather than left for you to find. The numbers are in §5 of the security statement. The part that matters most for you is the one thing concentration cannot touch: a signed delivery package verifies against a published key, with no account and none of our servers involved.


Learn more and verify