Security analysis for AI-generated code — fires the moment you paste, not while you type.

Most developers using AI assistants follow the same workflow: ask for code, review it quickly, paste it in. ShieldPrompt fits into that exact moment. The instant you paste, it silently analyses the code, highlights any vulnerable lines directly in the editor, and opens a findings panel explaining what's wrong and how to fix it — all before you've had time to run anything.

How it works

ShieldPrompt runs two analysis layers on every paste:

Static analysis matches known dangerous patterns on individual lines — fast, zero false positives on exact matches, covers the most common vulnerability classes across JavaScript, TypeScript, Python, and more.

Semantic analysis goes further. It builds a taint map across the entire pasted block, tracking user-controlled values as they flow through variable assignments, helper function calls, and into dangerous sinks — even when source and sink are separated by many lines. It also checks for structural problems: auth routes missing rate-limit middleware, upload handlers missing validation, truncated hash values used as identifiers.

Neither layer sends your code anywhere. Everything runs locally inside VS Code.

Why this instead of other security tools

Most security linters are designed for CI pipelines and whole-project scans. They require accounts, CLIs, language servers, or external services to work. ShieldPrompt has none of those dependencies — it installs as a single extension with no setup, works offline, and fires only on paste so it never slows your editor while you type.

It also detects whether pasted code was AI-generated, and labels findings accordingly — useful context when reviewing code from a model you don't fully trust.

Using it

Install the extension. Paste any block of code into a supported file. That's it.

If ShieldPrompt finds issues, a panel opens beside your editor listing each finding with its severity, the affected line, the rule that fired, and a plain-English explanation of the risk and the fix. Clicking a finding expands its detail. CVE and CWE references are clickable links.

The panel has four actions:

Re-scan File — runs a full analysis of the entire active file, not just the last paste. Before scanning it strips any existing ignore comments so previously suppressed lines are re-evaluated from scratch. Useful after you've made fixes and want a clean read.

AI Fix Suggestions — sends all current findings, along with the relevant lines of code for context, to the Anthropic API and displays fix proposals inline in the panel. To enable it, run ShieldPrompt: Set AI Fix API key from the command palette — your key is stored in the OS keychain (macOS Keychain, Windows Credential Manager, Linux libsecret), never in settings.json. If no key is set, the button falls back to copying a ready-made prompt to your clipboard so you can paste it into any AI assistant manually.

Scan Project — runs the full rule set against every file in the workspace and shows results grouped by file. Respects .gitignore.

Ignore on this line — available on each individual finding. Hides the finding in the panel without modifying your file. Re-scan File or pasting the same code again will bring the finding back, so nothing is silently lost.

Settings

Setting	Default	Description
`shieldprompt.enabled`	`true`	Enable or disable analysis on paste
`shieldprompt.minLines`	`3`	Minimum pasted lines to trigger analysis
`shieldprompt.suppressedRules`	`[]`	Rule IDs to suppress across the whole workspace
`shieldprompt.aiFixModel`	`claude-haiku-4-5-20251001`	Model used for AI Fix Suggestions
`shieldprompt.projectScanFileLimit`	`1000`	Maximum files the `Scan Project` command visits. Subsequent scans reuse a per-file mtime cache, so raising this is usually cheap
`shieldprompt.astLayer`	`true`	Drop findings whose position falls inside a comment or string literal (credential rules exempt). Template literals with `${}` are treated as code, so SQL injection via interpolation is still caught. Supported: JS/TS, Python
`shieldprompt.skipNoisyInTests`	`true`	In test / fixture / mock files, drop behaviour findings (eval, SQL concat, etc.) but keep credential findings. Detected via path patterns like `__tests__/`, `.test.`, `*_test.py`, `conftest.py`
`shieldprompt.minSeverity`	`"warning"`	`"warning"` (default) shows everything; `"critical"` hides warning-severity findings to reduce noise. Credential leaks always shown regardless

The Anthropic API key is not a setting — it's stored in the OS keychain via VSCode's SecretStorage. Set or clear it with the commands below.

Commands

ShieldPrompt: Open panel — Reopen the findings panel
ShieldPrompt: Toggle on/off — Disable temporarily without uninstalling
ShieldPrompt: Reset suppressed rules — Clear all workspace-level suppressions
ShieldPrompt: Scan Project — Run analysis across the full project
ShieldPrompt: Check for rule updates — Pull the latest rule definitions
ShieldPrompt: Set AI Fix API key — Store your Anthropic key in the OS keychain
ShieldPrompt: Clear AI Fix API key — Remove the stored key
ShieldPrompt: Clear project scan cache — Invalidate all cached per-file scan results for the current workspace

Privacy

All analysis runs locally inside the extension host process. Your code is never transmitted anywhere unless you use AI Fix Suggestions with an API key configured, in which case the flagged code context is sent to the Anthropic API. No account is required for any other feature. No telemetry.