PromptArmor BETA
Beta / experimental. Free while in beta. Feedback welcome. No guarantee of revenue or results.
Audits your AI prompts and agent config for injection exposure, leaked secrets, and missing guardrails.
Get the free beta by email Get a launch audit Send feedback How it works
What it does
Runs a deterministic, local-first, read-only scan over the AI layer of your project — system prompts, agent instruction files (CLAUDE.md, .cursorrules, .windsurfrules, copilot-instructions, AGENTS.md, *.prompt), prompt templates, tool/function definitions, and LLM call config — and returns a verdict (ARMORED / NEEDS HARDENING / EXPOSED), a 0-100 score, severity counts, redacted evidence for every finding, and a Claude-ready fix prompt per issue, in JSON, Markdown, and HTML. Runs entirely on your machine with zero runtime dependencies: no cloud, no account, no database, no billing, no telemetry.
Who it's for
Founders and AI-app builders shipping products with embedded prompts or autonomous agents who want a fast, honest reality check on the AI layer before untrusted input, real users, money, or reputation are on the line.
What it does NOT do
Reports prompt and agent readiness signals and evidence, not guarantees. It is not a security audit, a penetration test, or a legal or compliance check. It never executes your code, never reads or transmits real secrets (secret-looking values are redacted to [REDACTED] before they enter a finding), never deploys, bills, posts, emails, or uses paid services, and never crosses an owner manual gate. No guaranteed results, approval, or revenue.
How to try it
Request free beta access via the feedback form, then run it locally on Node: scan a bundled demo (node src/scan.mjs --demo bad or --demo good), point it at your own project, or open the local-only dashboard at 127.0.0.1 — no account or payment required.
What PromptArmor checks
- Secrets baked into prompts or agent files (API keys / tokens), with the value redacted to [REDACTED]
- Prompt-injection surface — untrusted input concatenated into a prompt with no delimiting or role separation
- Jailbreak / unrestricted-agent language (e.g. 'ignore previous instructions', 'you have no restrictions')
- Dangerous tool exposure — agent can reach shell / exec / delete / eval without an allow-list or confirmation gate
- Missing guardrail instructions — capabilities granted with no 'do not / never / ask before' boundaries
- Unbounded generation — LLM calls with no max_tokens / output cap (runaway latency and cost)
- Hardcoded or likely-outdated model ids scattered across the codebase
- Sensitive-PII collection instructions (SSNs, card numbers, and similar) inside prompts
- Auto-execute / auto-approve / no-confirmation agent flags on destructive actions
How it works
- Point PromptArmor at your project (or a bundled demo). It walks text and agent files read-only, skipping node_modules, .git, and build output — your source never leaves your machine.
- A deterministic check engine classifies AI-surface files and scores the result into ARMORED / NEEDS HARDENING / EXPOSED with a 0-100 score and severity counts, with all evidence redacted.
- Every finding includes a copy-ready Claude fix prompt that tells a Claude Code session exactly what to fix — and what not to do (no deploy, no real secrets, stop at any owner manual gate).
Known limitations
- Experimental beta — checks and scoring may change or miss things.
- Reports prompt/agent readiness signals and evidence only — no guaranteed results, approval, or revenue.
- Not a security audit, penetration test, or legal/compliance check; not yet reviewed for production use.
Changelog
- 2026-06-25 — Beta candidate created: local-first AI-layer audit engine (9 checks), CLI scan, JSON/MD/HTML reports, local-only dashboard, good/bad demo fixtures, and unit tests. Owner review required before promotion.
Try PromptArmor free while it's in beta
PromptArmor is free while in beta. Run it on your own prompts and agent config, then tell us whether the ARMORED / NEEDS HARDENING / EXPOSED verdict matched reality and what it missed — no account, no payment, and no guarantee of results.
Built by Everstead Labs LLC. Live on Shopify: LaunchLock and CatalogCare.