Connect the Safeguard MCP server to ChatGPT using the setup guide, sign in with OAuth, then ask about secrets findings in plain language. The assistant reads results Safeguard already produced. Secrets scanning itself runs in Safeguard, and a new workspace starts with only a few read-only tools enabled.
What does Safeguard's secrets scanning look at?
Per the Safeguard docs, it looks for leaked credentials, API keys and tokens in source code at HEAD, Git history, container image layers, packaged artifacts, build logs, vendor SBOMs and model weights. Detection works in layers: issuer-specific patterns for more than 200 providers, generic patterns such as private keys and JWTs, and entropy checks for high-entropy strings.
For issuer-specific matches it makes a low-privilege call to the issuer to see whether the credential is live. Findings it cannot verify are still raised and marked unverified. Severity follows: critical for a verified high-privilege credential, then high, medium and low for generic or entropy-based matches.
How do you connect ChatGPT?
The MCP page at safeguard.sh/mcp links a setup guide for ChatGPT, and the docs describe the general flow. Add the server URL in your assistant, and you are sent to Safeguard's sign-in page at app.safeguard.sh, where you check the app name and select Authorize access. There are no keys to copy. A fallback MCP key exists for clients that send a fixed header, created in Settings, then Developer, and shown once. Follow the ChatGPT guide for the exact steps, because we have not reproduced its screens here.
What can you ask once connected?
The docs say visibility is set per tenant. A new workspace starts with six read-only tools for browsing products, projects, vulnerabilities and findings, and an administrator enables others per tool under Feature Flags. So your first questions should be read questions:
- Which projects have critical findings?
- Show findings flagged as secrets, newest first.
- Which of these are verified, and which are unverified?
If a question needs a tool your workspace has not enabled, the assistant will say so. Ask your admin.
What can the chat not do?
The docs state there is no CLI command for a Git pre-commit hook or for listing, revoking or rotating secret findings. Those tasks use the IDE extension, which blocks secrets before push, or the Secrets dashboard. Revocation steps for verified findings are a playbook there, with optional approvals. Treat the chat as a way to look, not as your revocation tool.
What to do
- Open the ChatGPT setup guide from safeguard.sh/mcp and add the server.
- Authorize with OAuth, checking the app name and host on the sign-in page.
- Ask for projects with secrets findings and sort by verified status.
- For each verified finding, revoke the credential at the issuer first, then rotate it.
- Remove the secret from Git history only after rotation, if your policy requires it.
- Ask your admin whether to enable more tools, and review the list.
- Turn on the IDE extension so new secrets are blocked before push.
What does a good first session look like?
Start by asking which projects have the most open critical findings, and filter to secrets. Pick the top project and ask for the list of findings with their severity and verification status.
Read the verified ones first. A verified critical finding means a live, high privilege credential was seen in your data. Open your cloud or provider console and revoke it at the issuer now. Do not wait for the code change to land.
Next, ask where the secret appeared: the file, the commit, a build log or an image layer. That tells you what else needs cleaning. A secret in a container layer, for example, may be baked into images that are already deployed.
Finally, ask for the unverified matches and sort them by issuer. Many will be test values. Use the allowlist options in Safeguard for real false positives, and record why, since the docs say allowlist changes are audited.
Quick check
Does ChatGPT see my source code?
The assistant sees what the MCP tools return, such as findings and project data, based on the tools your admin enabled. Review that list before connecting.
Should I revoke or delete the commit first?
Revoke and rotate first. A secret that was public for even a short time must be treated as exposed.
Sources
Last reviewed: 18 October 2026.
