BUILD: secureai scan

Static analysis for prompts, tool definitions, and templates. It runs offline, and prompt content stays on the machine, so you can use it on every commit.

Terminal

secureai scan <target> [flags]

<target> is a file, a directory, or a glob. Directories are walked recursively.

Common uses

Terminal

# Scan a directory of prompts
secureai scan ./prompts

# Scan a single file
secureai scan prompts/system_prompt.md

# Fail the command if anything High or worse is found
secureai scan ./prompts --fail-on high

# Machine-readable output for a pipeline
secureai scan ./prompts --format json > findings.json

# Only run the prompt-injection rules
secureai scan ./prompts --rules R-01,R-02,R-03

# Write a full HTML assessment alongside the terminal output
secureai scan ./prompts --report assessment.html

Flags

Flag	Default	Description
`--config`	`secureai.config.yaml`	Path to the config file. See Configuration.
`--fail-on`	none	Exit non-zero when a finding reaches this severity: `informational\|low\|medium\|high\|critical`. Overrides `fail_on` in the config file.
`--format`	`terminal`	Output format: `terminal\|json\|jsonl`.
`--report`	none	Also write a full assessment report to this path (`.html`). See Reports.
`--rules`	all	Comma-separated rule IDs to run, overriding the config file.
`--no-bus`	off	Do not emit a SecurityEvent to the local Intelligence Bus.

The report is written before the gate

--report produces its file even when --fail-on trips and the command exits 1. A failing build still leaves you the artefact that explains why.

What it looks for

Nine rules, R-01 through R-09, each mapped to an OWASP LLM Top 10 category.

Rule	Name	OWASP
`R-01`	Prompt Injection Risk	LLM01
`R-02`	Instruction Override	LLM01
`R-03`	Unsafe Role Assignment	LLM01
`R-04`	Missing Refusal Logic	LLM06
`R-05`	Over-Permissive Tools	LLM06
`R-06`	Sensitive Data Exposure	LLM02
`R-07`	Unrestricted Output Format	LLM05
`R-08`	Missing Context Boundaries	LLM04
`R-09`	Ambiguous Capability Scope	LLM09

Each rule's detection logic and how to fix what it flags is on Scan rules.

Reading the output

Terminal

BUILD: static risk scan
14 files scanned · 18 findings · score 64/100

CRITICAL  R-02  Instruction Override
          prompts/system_prompt.md:12-18  ·  LLM01 Prompt Injection
          The prompt contains no precedence rule, so user text asking the
          model to ignore prior instructions is honoured.

HIGH      R-06  Sensitive Data Exposure
          prompts/support.md:31  ·  LLM02 Sensitive Information Disclosure

gate: fail-on=high → FAILED

Score is out of 100 and higher is safer. It weights findings by severity across the files scanned.
Location is file:line or file:start-end, so editors and CI annotations can link straight to it.
Gate is only printed when --fail-on or a config fail_on is in effect.

Pre-commit hook

Install a git hook so a scan runs before every commit in the current repository:

Terminal

secureai scan install-hook

What gets scanned

Prompt files (.md, .txt, .prompt), tool and agent configuration (.yaml, .yml, .json), and templates. Files it cannot parse are skipped rather than failing the run.

Nothing leaves the machine

Analysis is local and pattern-based. There is no model call or prompt upload. Only the resulting findings are sent to your dashboard, and --no-upload suppresses even that.