Guard — catch AI risk before it ships

Guard is authoring-time protection: catch AI-written and AI-adjacent risk in the editor, PR, and CI before it reaches production. Runtime agent governance lives in Govern — same Monitor Mode discipline, different control point.

Added value

Without Guard	With Guard
LLM-speed code merges unreviewed	Deterministic findings on secrets, insecure patterns, IaC, typosquats
CI fails on whole-file noise	Diff-aware scan judges added lines only
Separate local vs CI tools	One detector floor: CLI = IDE = Action = API
Day-one break-the-build	Monitor Mode → enforce when signal is trusted

What it catches

Category	Examples
Secrets & PII	AWS keys, tokens, private keys, Luhn-valid credit cards
Insecure code (LLM-typical)	eval/exec, command injection, disabled TLS, weak crypto, dangerouslySetInnerHTML, broad CORS, unsafe deserialization
IaC exposure	public S3, iam:*, 0.0.0.0/0, public databases, broad AI IAM (bedrock:*)
Hallucinated / typosquatted deps	packages an LLM invented, or a near-miss of a popular one
Agent / MCP risk at commit	Unregistered agents, risky configs, new tool permissions (feeds Govern)

Full reference with examples: What Unveilr detects.

Local scan (developer)

unveilr scan                 # observe: reports, exits 0
unveilr scan --json          # IDE / scripts

VS Code / Cursor extension: Problems panel from the same --json output — Developer experience.

Screenshot to capture — Code Findings

Console path: /findings
Capture: list sorted by risk score / blast-radius; open one finding with path and severity.
Story it tells: “AppSec triages AI-SDLC risk the way they triage other vulns.”

Screenshot to capture — Remediation

Console path: /remediation and the Findings drawer
On a connected GitHub repo with App contents + pull_requests write accepted on the installation, Open fix PR opens a remediation branch/PR. Ticket actions remain available without write.

CI gate — Monitor Mode → Enforce

unveilr scan --mode enforce --fail-on high

Exit codes: observe → always 0; enforce → 1 when a finding is at or above --fail-on (info | low | medium | high | critical, default high).

GitHub Action (recommended):

- uses: unveilr/scan-action@v1
  with:
    mode: observe
    fail-on: high

Screenshot to capture — PR check

Where: GitHub PR checks + inline annotation on an added line.
Story it tells: “We block (or advise on) the risk this PR introduced.”

PR checks — judge only what a change introduces

git diff origin/main...HEAD > /tmp/pr.diff
unveilr scan --diff /tmp/pr.diff --mode enforce --fail-on high

The diff scanner keeps only findings on added lines. See Integrations for the GitHub PR-check API and Actions setup.

Authoring-time ↔ runtime

Authoring-time Guard secures the code assistants write; Govern secures the actions agents take. Unregistered agents discovered at scan time become high-severity findings that can fail the PR that would ship them — then Govern owns the live credential and tool scope.

Where to next

• Govern — Gateway, scopes, approvals, sessions.
• Prove — triage, blast-radius, evidence, compliance.
• Integrations — wire Guard into CI and tickets.