Skip to content
Prompt Guardian
Start free

Prompt governance for agents that act

No prompt change ships until it survives the attack.

Prompt Guardian puts every system prompt change through three gates: a static safety scan, an adversarial evaluation by a second model, and your explicit approval. Red-team it with attack packs and hand security an OWASP report built from real results.

Guard your first prompt free See pricing
  • Free plan with no card. Sign in with Google in one click.
  • Nothing overwrites your active prompt without a human approval.
  • Every run keeps its scores, patch, rationale and the reasons it was blocked.
A refinement run awaiting human approval, with before and after evaluations and the behavioral diff
A refinement run awaiting human approval, with before and after evaluations and the behavioral diff. Real product screen.

The risk you can't unit test

One edited sentence can turn a careful agent into a compliant one.

System prompts drive agents that send messages, submit forms and move data. A well-meant edit can drop a guardrail, and a single line in a retrieved document can override it. Reviewing the diff by eye does not show what the model will do; attacking it does.

#1

Prompt injection (LLM01) is the top risk in the OWASP Top 10 for LLM Applications 2025, for the second edition in a row.

Source: OWASP Top 10 for LLM Applications

Three gates, then you decide

  1. 01

    Store the prompt

    Paste or upload your system prompt. Prompt Guardian versions it and runs the static safety scan as you type.

  2. 02

    Refine

    A refiner model proposes a patch from verified issues or your feedback. Truncated or unsafe candidates are rejected on the spot.

  3. 03

    Attack

    A second model scores the candidate against critical rules. One unmet rule fails the run, whatever the refiner claimed.

  4. 04

    Approve

    Cleared candidates wait on the Runs page with before and after scores. Promotion is a deliberate two-step confirmation with a backup.

Evidence, not opinions, for every prompt you ship

Each feature produces something you can show a reviewer: a score, a breached case, a mapped OWASP category, a pass or fail verdict.

Human-gated refinement

See exactly what a change fixes and what it breaks

Every run records the before and after evaluations and a behavioral diff: which rule violations were cleared and which new ones appeared, so you approve on facts.

  • Refiner and adversary are different models you choose
  • Static scan blocks auto-submit, credential and bypass language
  • Each promotion snapshots the prompt it replaced
A refinement run awaiting human approval, with before and after evaluations and the behavioral diff

Attack packsPro

Red-team the prompt with ready-made attacks

Run injection, exfiltration, obfuscation, role-play and system-prompt extraction cases against the active prompt or a pinned version. A case passes only when the attack is defended.

  • Baseline direct-injection pack free; five Pro packs with 20 more cases
  • Each case mapped to OWASP LLM ids
  • Deterministic leak checks run after the judge
Attack pack results: ten attacks defended and one indirect injection breached, each mapped to OWASP LLM ids

Compliance reportPro

An OWASP LLM Top 10 report your security team can read

Turn stored attack runs, suite results and the static scan into a report that maps every finding to LLM01–LLM10 with evidence and remediation, and says plainly what was not assessed.

  • Optional AI executive summary
  • Export as Markdown, HTML or PDF
  • Regression guard compares two versions and exports a CI verdict
An OWASP LLM Top 10 (2025) compliance report with pass, fail and not-assessed categories

Prompt workspace

Your system prompt, versioned like code

A live Markdown editor with an outline, scan status as you type, and a version history you can roll back, plus your own regression cases that every refinement must keep passing.

  • Regression suite of your own test cases
  • Releases ledger with rollback
  • Compare models on the same prompt
The active system prompt in a live Markdown workspace with outline and scan status

For the people who sign off on what the agent is allowed to do

Security engineers attack the prompt, the team reviews the evidence together, and compliance gets a report instead of a promise.

A security engineer studies code late at night under a warm desk lamp
Security engineers run attack packs against each candidate before it ships.
Three engineers sketch a threat model on a glass whiteboard
Teams review breached cases and agree on the fix in one place.
A compliance officer signs off a printed security report with a young engineer
Compliance signs off on an OWASP-mapped report built from stored results.

Photos are AI-generated illustrations, not customers.

Check the prompt from anywhere

The editor, scan status and run history work on a phone, so a review never waits for a laptop.

Prompt Guardian's prompt editor on a phone

a DevSpeak Labs product

One plan, four apps.

Prompt Guardian is one of four DevSpeak Utilities apps for the inputs to AI agents: specs, system prompts, technical text and prompt tests. Each app is free to use, and one Utilities Pro plan ($5.99/month or $59.99/year) unlocks Pro in all four.

  • Prompt Guardian's refinement dashboard: a support-agent system prompt in the editor with a clean static scan

    Prompt Guardian

    Gated prompt governance

    You are here

    Attack, score and gate every system prompt change before it ships.

    Pro adds in Prompt Guardian

    • Advanced attack packs
    • OWASP LLM Top 10 (2025) report
    • Regression guard for CI
  • SpecStudio's Studio: a travel-booking request with allowed domains, constraints and success criteria, ready to compile into a spec

    SpecStudio

    Agent task specs

    Turn a rough request into a safety-gated spec a browser agent can follow.

    Pro adds in SpecStudio

    • Spec risk audit (OWASP LLM Top 10)
    • Export to AGENTS.md, CLAUDE.md and Cursor rules
    • Live page grounding
    • Pro template pack
    Open SpecStudio
  • Dev Text Forge's workbench: an informal note about a deploy script in the source pane, next to the optimized output pane and quality gates

    Dev Text Forge

    Dev text optimizer

    Rewrite developer text into clear technical English that keeps every token.

    Pro adds in Dev Text Forge

    • Agent context mode
    • Token diet report
    • Batch optimize
    Open Dev Text Forge
  • PromptForge's tool directory: prompt-engineering tools such as System Prompt Maker, Prompt Optimizer and Image to Prompt

    PromptForge

    Prompt engineering toolbox

    A prompt-engineering toolbox: generators, a linter, tests and a workspace.

    Pro adds in PromptForge

    • Prompt linter
    • One-click prompt hardening
    • Assertion-based prompt tests
    Open PromptForge

Free to start. One price for Pro everywhere.

Use Prompt Guardian free for as long as you like. When you want the Pro features, one plan covers all four utilities.

Prices in USD, taxes included. One plan for all four apps.

Free

Everything Prompt Guardian does today. No card, no trial clock.

$0forever

Sign in with Google or email.

  • Static safety scan and versioned prompt workspace
  • Refinement runs with adversarial evaluation
  • Human-gated promotion with backups and rollback
  • Your own regression test cases
  • Baseline direct-injection attack pack
Start free
Recommended

Utilities Pro Pro

Pro features in Prompt Guardian and in the other three utilities.

$5.00per month

Billed $59.99 once a year. Save 17% versus monthly.

7-day Pro trial, no card required

Sign in with Google and it starts by itself: every Pro feature in all four apps, with up to 30 Pro actions a day (100 in total) and 3 premium-model runs a day (5 in total). Then you're back on Free automatically; nothing to cancel.

One trial per email. Not available to emails that already had Utilities Pro, and new trials are limited each day.

  • Everything in Free
  • Advanced attack packs
  • OWASP LLM Top 10 (2025) report
  • Regression guard for CI
  • Premium models
  • Pro editor: AI assist, cloud history, HTML/PDF export, token budget

Pro in all four apps

  • Prompt GuardianThis app
  • SpecStudioAgent task specs
  • Dev Text ForgeDev text optimizer
  • PromptForgePrompt engineering toolbox
Free and Utilities Pro compared, app by app
FeatureFreePro
In Prompt Guardian
Static safety scan and versioned prompt workspace
Refinement runs with adversarial evaluation
Human-gated promotion with backups and rollback
Your own regression test cases
Baseline direct-injection attack pack
Advanced attack packsRed-team your prompt with indirect injection through documents and tool output, Markdown exfiltration, encoding obfuscation, multi-turn role-play jailbreaks and system-prompt extraction.
OWASP LLM Top 10 (2025) reportTurn attack, suite and evaluation results into a stakeholder report mapped to LLM01–LLM10, with evidence, remediation and Markdown, HTML or PDF export.
Regression guard for CICompare two prompt versions or runs, flag attacks that newly succeed and score drops past your threshold, and export a pass/fail JSON verdict for CI.
Premium modelsThe large models in every model picker: Claude Opus 5.5, Claude Sonnet 5.5, Mistral Large 4 and Mistral Medium 3.5. Claude Haiku 5.5 (the default), Mistral Medium 3.1, Devstral 2, DeepSeek and the other fast models stay free for everyone.
Pro editor (every app)
AI writing assist in the editorRewrite, tighten, fix, expand or turn any selection into a checklist.
Cloud version historyEvery save kept in the cloud, with a side-by-side diff and one-click restore into your draft.
HTML and PDF exportShare polished prompts with stakeholders who never open a Markdown file.
Token and context budgetSee how much of each model's context window your system prompt uses.
In SpecStudio
SpecStudio: everything free todayTurn a rough request into a safety-gated spec a browser agent can follow.
Spec risk audit (OWASP LLM Top 10)
Export to AGENTS.md, CLAUDE.md and Cursor rules
Live page grounding
Pro template pack
In Dev Text Forge
Dev Text Forge: everything free todayRewrite developer text into clear technical English that keeps every token.
Agent context mode
Token diet report
Batch optimize
In PromptForge
PromptForge: everything free todayA prompt-engineering toolbox: generators, a linter, tests and a workspace.
Prompt linter
One-click prompt hardening
Assertion-based prompt tests
  • Free stays free. Pro only adds; nothing you use today moves behind the paywall.
  • Cancel anytime from Settings → Billing. Pro stays active until the end of the period you paid for.
  • One subscription, four apps. In another app, sign in with Google using the same email and Pro is already on.
  • Secure checkout by Dodo Payments, merchant of record. It handles sales tax and VAT; prices include them.

Questions, answered

Can Prompt Guardian change my production prompt on its own?

No. Runs never write to the active prompt. A candidate that clears every automated gate waits for your approval, and promotion keeps a backup of the prompt it replaced.

Is the free plan really free?

Yes. The scan, refinement runs, human-gated promotion, your regression cases and the baseline attack pack are free with no card and no trial period.

What does Utilities Pro add here?

Five advanced attack packs, the OWASP LLM Top 10 (2025) report with export, the regression guard with a CI verdict, the Pro editor tools, and the premium models: Claude Sonnet 5.5, Claude Opus 5.5, Mistral Large 4 and Mistral Medium 3.5. The same plan unlocks Pro in the other three utilities.

Which models can I use?

Claude Haiku 5.5 by default for the refiner and the suite, with Mistral Medium 3.1 via OpenRouter (processed in the EU) as the automatic fallback. Mistral Medium 3.1 (via OpenRouter, processed in the EU) is the default adversarial evaluator and judge, so a candidate is not scored by the model that wrote it. Free: Claude Haiku 5.5, Mistral Medium 3.1, Devstral 2, and DeepSeek, Xiaomi MiMo and Zhipu GLM, which run only when you pick them and are processed in China. Utilities Pro adds Claude Sonnet 5.5, Claude Opus 5.5, Mistral Large 4 and Mistral Medium 3.5.

How do I use the regression guard in CI?

Export the guard's JSON verdict and check it in your pipeline, for example with jq -e '.exitCode == 0' guard.json. There is no CI API endpoint yet; the JSON comes from the app.

Can I cancel Utilities Pro?

Anytime, from Settings → Billing. You keep Pro until the end of the period you paid for, then the free plan continues.

Ship prompt changes the way you ship code: reviewed, tested, approved.

Start free with the three gates. Upgrade for attack packs and the OWASP report, in this app and the other three.

Guard a prompt free