Security model
Trust boundaries, what each component can see, how PAM credentials stay inside your network and out of logs and uploads, and the controls you have over what a runner may read.
Security model
Trust boundaries
| Component | Holds a PAM value? | Holds a PAM connection secret? | Holds a stored vault value? |
|---|---|---|---|
| Safeguard platform (vault, scan targets, findings) | Never | Never | Yes, encrypted with AES-256-GCM. Credentials with a policy and all their versions use a key of your organization alone |
| Safeguard cloud scanner | Never (PAM references are refused) | Never | During a cloud scan of a stored entry |
| Your runner | In memory, during the scan that read it | Yes, in its provider file and the files it points to | During a scan that uses a stored entry |
| Your PAM | Yes | Its side of the identity | No |
A person with access to the Safeguard platform cannot read a PAM value through it: a PAM entry has nothing to reveal, and its history shows only outcomes.
Credentials kept in Safeguard
For a credential in the Safeguard vault with a policy on:
- Encryption. The value and every version are encrypted with AES-256-GCM under a data key that belongs to your organization, which is itself wrapped by the platform key. A ciphertext is bound to the organization, credential and version it was written for, so it cannot be decrypted if copied elsewhere. See Per-organization encryption keys.
- Write-only. No screen, API call or CLI command returns the value. Only a check-out by a scan that is granted the credential, bound to the target and, if required, approved, returns it, with
Cache-Control: no-store. Revealing needs the REVEAL grant and is audited. - Least privilege. No role implies REVEAL, and
vault-admindoes not imply USE. An AI agent can use a credential it was granted and can never manage, approve or reveal one. - Short leases. A check-out ends when the scan checks in, or at the lease's end. A lease can be revoked.
- Separation of duty. Nobody approves their own request, and approving, revealing and changing grants can require a recent sign-in.
- Evidence. Every check-out, check-in, request, decision, version, rotation, grant change and export is in the audit log, which holds no value.
- Rotation without exposure. The runner makes the new secret on its own machine, the vault stores it encrypted before it is used, and a failed rotation rolls the target back. Rotation settings are refused if they name a secret.
Controls you hold
- The PAM identity. Safeguard reads only what the identity you created may read. Give it retrieve rights on the scan accounts only and, where your PAM supports it, bind it to the runner's address or host.
allowed_refson the runner. Each provider can list the references the runner may read, as patterns. Anything else is refused on the runner withNOT_ALLOWED, before any request to the PAM, whatever a scan target asks for.- Which runner. DAST, API and host scans are addressed to the runner chosen on the target and require its declared capability. No other runner can claim them.
- The provider file. It lives only on the runner, must be private to the runner's account (refused otherwise on macOS and Linux), and refuses inline secrets: keys and tokens are
env://orfile://pointers, read at the moment of use. Certificates and keys are file paths, never registered in the app. - Turning it off. Removing a provider from the file, or the file itself, stops all reads from that PAM at the next scan.
Data in transit
- Runner to PAM: TLS 1.2 or later with certificate verification, your CA through
ca_cert, mutual TLS throughclient_certandclient_key. A plainhttp://address is refused except for the runner machine itself, unlessallow_http: trueis set, which logs a warning on every request.tls_insecure_skip_verifyexists for short tests and logs a warning on every request. - Redirects from the PAM are not followed, so a token or key is never sent to another address.
- Runner to platform: outbound HTTPS only. Nothing connects to the runner.
Values on the runner
- A value read from a PAM is held in memory for the scan that read it, cached so one scan reads each reference once (a DAST re-login reads again), and zeroed and forgotten when the scan ends. Go strings derived from it cannot be zeroed and are left to the garbage collector.
- Every value of 6 characters or more is registered with a redactor that replaces it with
[redacted]in log output, error messages, job errors, findings, evidence, uploads and crash output, including its JSON-escaped, URL-encoded and HTML-escaped forms. Shorter values are not redacted. Base64 forms (for example inside an HTTP Basic header the target logs) are not. - Reading a reference never puts its value on a command line or in a process environment; a repository token read from a PAM reaches git the same way a typed token does. Scan job spool files on disk hold the reference strings, never the values.
- What the provider holds is released when the scan ends: a BeyondTrust request is checked in, HashiCorp Vault dynamic-secret leases and login tokens are revoked.
Failure behaviour
- If a reference cannot be read, the scan does not start. Safeguard never sends the reference text as a credential and never scans with an empty one.
- An on-prem scan never falls back to the cloud.
- Retries happen only for transient failures (connection, timeout, 429, 5xx), never for refusals, so a wrong key does not lock an account through repeated attempts.
Accountability
Every read and test is recorded on the platform, in the runner's local audit file and, where it records a reason, in your PAM's audit, with the scan id and runner name to join them. See Audit.
Operating checklist
- A dedicated OS account for the runner; provider file, key files and certificates owned by it, mode 600.
- A dedicated PAM identity for Safeguard, least privilege, no approval step, bound to the runner host where possible.
allowed_refsset for each provider.- The local audit file collected into your SIEM.
NO_PROXYset for the PAM host if the runner uses a proxy.- Rotate the PAM identity's own key on your usual schedule; update the key file on the runner, no restart needed.
Audit
What is recorded for every credential read, check-out, approval, rotation and test, on the platform, in the runner's local audit file and in your PAM's own audit, and how to filter and export the tenant-wide log.
Troubleshooting
Error codes for credential reads, what each means and how to fix it, and the common problems with PAM entries, runners, DAST, API and host scans.