CLI overview

One binary, four phases. Each phase is a top-level command that can be run on its own, locally or in CI.

Terminal

secureai [command] [flags]

Available Commands:
  proxy       Expose a configured text model as OpenAI Chat Completions
  run         Run BUILD, TEST, VALIDATE, and RUNTIME from one manifest
  scan        Phase 1 BUILD: statically scan prompts and tool configs for risk
  test        Phase 2 TEST: offensive prompt-injection / jailbreak testing
  validate    Phase 3 VALIDATE: RAG security analysis and guardrail checking
  runtime     Phase 4 RUNTIME: context access control, behaviour fingerprinting,
              capability enforcement
  conformance Replay the CoreLayer CSV suites through each phase and score accuracy
  login       Authenticate the CLI via your browser (OAuth loopback + PKCE)
  logout      Clear stored CLI credentials
  whoami      Show the currently authenticated CLI user

Model accessUniversal proxyTranslate common provider APIs or a custom JSON contract into one test endpoint.LifecycleRun all phasesExecute every phase in order from a versioned manifest.Phase 1CSA: scanStatic analysis of prompts, tool configs, and templates. No model calls.Phase 2ART: adversarialFire the payload corpus at a live endpoint and measure attack success rate.Phase 3DCA: deploymentScore a RAG pipeline or system prompt out of 100 before release.Phase 4RTD: enforcementContext access policy, behavioural drift, and capability boundaries.

Global flags

These apply to every command.

Flag	Default	Description
`--server`	`$SECUREAI_SERVER`	Control-plane URL used for result upload and payload fetch.
`--no-upload`	off	Run entirely locally. Results are printed but not sent to your organisation.
`-v, --version`	-	Print the CLI version.
`-h, --help`	-	Help for any command or subcommand.

Results upload by default

Every phase reports to your organisation's dashboard so security work is visible without anyone forwarding terminal output. Use --no-upload when you are experimenting and do not want the run recorded.

Output formats

Every phase accepts --format. Use terminal when a human is reading and json when a machine is.

Format	Available on	Use it for
`terminal`	all phases	Human-readable, coloured, grouped by severity. The default.
`json`	all phases	One object per run. Parse with jq, store as a build artefact.
`jsonl`	`scan`	One finding per line. Streams into log pipelines cleanly.

Failing a build

Each phase has a gate flag that decides the exit code. This is what turns CoreLayer into a check rather than a report.

Phase	Gate flag	Meaning
`scan`	`--fail-on`	Exit 1 if any finding is at or above this severity.
`test`	`--fail-asr`	Exit 1 if attack success rate exceeds this percentage.
`validate`	`--fail-below`	Exit 1 if the deployment score falls below this number.

Exit 0 means the gate passed, 1 means it failed, and 2 means the CLI itself errored: bad flags, an unreachable server, or an unreadable file. A gate failure and a broken pipeline are different problems, so they get different codes. See Scores & exit codes.

Commands that need authentication

All of them. Every phase requires a session, as do run, proxy, and conformance. Without one the CLI exits 2 before doing any work. Only --version, --help, and login / logout / whoami run logged out, so you can install the binary and check it before you have an account.

The session check is a local one, which is why scan and validate still make no network calls and remain usable in a pre-commit hook or an air-gapped build. test is the exception in the other direction: its attack payloads are not in the binary at all but fetched from the control plane per run, so that phase is authenticated by the server rather than by the CLI.

Use --no-upload to keep results off the dashboard. It suppresses the upload, not the session requirement. Provider credentials are separate and belong in environment variables referenced by the model proxy config. See Authentication.