sarix 0.4.0

Semantic security verification CLI for scanner alerts and AI-written code

Sarix

Security CLI for scanner alerts, code review, and CI reports.

Sarix started as a direct code scanner. Now the main lane is stronger: take alerts from Semgrep, CodeQL, or another SARIF tool, open the matching source files, collect evidence, and decide what deserves attention.

It still scans files and folders directly. The bigger value is verification: less noise, clearer proof, and reports that can go back into CI.

What It Is Now

Sarix has three useful paths:

• guided terminal UI for people who do not want to remember commands
• direct scan for one file, a folder, or a git diff
• SARIF verify for scanner alerts that need real source context

The verify path is the important one.

Scanners are good at breadth. They find suspicious lines fast. Sarix checks what is behind the alert:

• where the input comes from
• which sink or trust boundary it reaches
• what guard is missing
• why the verdict is exploitable, likely exploitable, uncertain, or dismissed
• what fix would actually reduce risk

No evidence, no confident verdict.

Quick Start

Guided UI:

pip install sarix
sarix

From the repo:

python -m sarix

Pick a path in the menu:

• scan a file
• scan a folder
• review a git diff
• verify a SARIF file
• build a proof bundle
• run judge and skeptic mode

Direct commands still work:

sarix security src/app.py
sarix scan src --task security --output sarif > sarix.sarif
sarix diff --task security

Local model:

ollama serve
sarix security src/app.py --provider ollama

Verify Scanner Alerts

Run your scanner first:

semgrep scan --config auto --sarif > semgrep.sarif

Then let Sarix import the alert and attach local source context:

sarix verify semgrep.sarif --source . --output markdown

Useful verify modes:

sarix verify semgrep.sarif --source . --preview-context
sarix verify semgrep.sarif --source . --fail-on likely_exploitable
sarix verify semgrep.sarif --source . --judge --skeptic --profile auto --provider openai
sarix verify semgrep.sarif --source . --artifact-dir .sarix-proof

Plain import mode is conservative. It keeps scanner alerts as uncertain. Judge mode can promote, downgrade, or dismiss alerts. Skeptic mode checks serious verdicts again before CI has to trust them.

Typical summary:

Blocked: 0 exploitable issue(s)
Likely exploitable: 0
Needs review: 42
Dismissed: 0

Try the local fixture:

sarix verify examples/semgrep-verify/semgrep.sarif \
  --source examples/semgrep-verify/project \
  --output markdown

Framework fixtures are in examples/framework-profiles.

Evidence Format

Sarix uses structured verdicts:

• exploitable
• likely_exploitable
• uncertain
• probably_false_positive
• not_exploitable

Example:

### CS-AUTH-001: Tenant isolation bypass

- Verdict: `exploitable`
- Severity: `high`
- Confidence: `high`
- Exploitability: `91/100`
- Location: `api/projects.py:88`
- CWE: `CWE-862`

#### Evidence

- Source: `request.path_params["org_id"]`
- Sink: `Project.query.filter_by(org_id=org_id)`
- Missing guard: no membership check before project lookup

#### Evidence path

1. `api/projects.py:82` - route accepts org_id from the request path
2. `api/projects.py:88` - query trusts org_id before checking membership

Commands

Core:

• sarix
• sarix ui
• sarix security <file>
• sarix scan <dir> --task security
• sarix diff --task security
• sarix verify <scanner.sarif> --source .
• sarix benchmark

Secondary:

• sarix review <file>
• sarix bugs <file>
• sarix docs <file>
• sarix explain <file>
• sarix refactor <file>

Providers

• OpenAI: OPENAI_API_KEY
• Anthropic: ANTHROPIC_API_KEY
• Google Vertex AI: GOOGLE_CLOUD_PROJECT and ADC
• Ollama: local ollama serve
• OpenAI-compatible: custom label from sarix config

OpenAI-compatible presets include OpenRouter, Groq, Together AI, Mistral, xAI, Fireworks, DeepSeek, Perplexity, Cerebras, Cohere, and Azure AI Foundry.

Output

sarix security app.py --output markdown
sarix security app.py --output json
sarix security app.py --output sarif > sarix.sarif
sarix verify semgrep.sarif --source . --output sarif > verified.sarif

SARIF can be uploaded to GitHub code scanning.

Privacy

Sarix does not need a hosted account or repo connection.

• Ollama keeps analysis local.
• BYOK providers use your own key.
• Project config cannot set api_key, base_url, or default_provider.
• Project config discovery is restricted to $HOME.
• Large files can be compressed into code maps before prompting.

Cloud providers still receive the selected code context. Use Ollama when code must stay on the machine.

Benchmarks

The built-in benchmark is a smoke test: 10 vulnerable Python cases and 2 clean false-positive traps. It is useful for checking provider behavior and prompt drift. It is not a public claim that Sarix is better than another scanner.

sarix benchmark --models gpt-5.4 llama3
sarix benchmark --json > benchmark-results.json

Public benchmark claims need the exact cases, commands, expected verdicts, raw results, model, provider, and run date.

GitHub Action

Verify scanner SARIF:

- run: |
    python -m pip install semgrep
    semgrep scan --config auto --sarif --output semgrep.sarif

- uses: AvixoSec/sarix@v0.3.1
  with:
    mode: verify
    path: .
    sarif-input: semgrep.sarif
    output: sarif
    fail-on: exploitable
    judge: "true"
    skeptic: "true"
    profile: auto

Direct scan:

- uses: AvixoSec/sarix@v0.3.1
  with:
    provider: openai
    api-key: ${{ secrets.OPENAI_API_KEY }}
    task: security
    path: .
    output: sarif

Development

git clone https://github.com/AvixoSec/sarix.git
cd sarix
pip install -e ".[dev]"
pytest tests -v
ruff check .

License

MIT. See LICENSE.