# Website policy.

> One place to understand our product scope, published results, account rules and payment terms.

Official source: https://warhoundai.com/website-policy/
Updated: 2026-11-06

## What Mini Hound reviews

Mini Hound reads public pages and code on your verified website, or source excerpts you provide. It explains possible issues and saves source evidence. It cannot obtain private backend code or enter private accounts from a public URL.

Source observations are not automatically verified exploits. Mini Hound does not run active attacks, high-volume requests, destructive tests or automatic production changes. Mini Hound v1 prioritizes speed; v1 Pro prioritizes quality. Both have the same scoped source, isolated browser and synthetic role-check tools. BloodHound’s broader assessments, continuous monitoring and automatic code changes remain in development.

AI answers and suggested code changes can be incomplete or incorrect. Check the evidence, test changes and verify the result in your application. A review does not guarantee that a website is secure or that every issue has been found.

- [Current product availability](https://warhoundai.com/product-status/)
- [Security and data handling](https://warhoundai.com/security/)

## Account access and website permission

Verify your email, activate Free or Plus, then verify the exact HTTPS website you want to review. Other origins, subdomains and ports are not automatically included. Standard website verification lasts 30 days and proof codes expire after 24 hours.

Use only systems you own or have permission to assess. A public URL does not grant permission. Operator-managed demo exceptions can be time-limited and remain restricted to the approved scope. Server-side safeguards and plan limits apply regardless of wording in a prompt.

- [Acceptable use policy](https://warhoundai.com/acceptable-use/)

## Your information and evidence

Chat messages and relevant collected source are processed by the configured AI provider. Cloud records, source artifacts and recovery checkpoints are private to the authorized account. Do not submit live credentials or unnecessary personal information.

Read-only refers to access to your website. Chats, findings and recovery records are saved in your WarHound account. An administrator’s evidence review is separate from an automated exploit test. Provider processing, retention, deletion and optional voice-input details are explained in the privacy notice.

- [Privacy notice](https://warhoundai.com/privacy/)

## Prices, fees and usage

Free is $0/month with $0.50 in credits each UTC calendar month and Mini Hound v1. No activation payment is required. Plus is $4.99 USD per purchased month with $3 in credits and both models. Renewal remains manual. Unspent credits carry over.

Each agent is independently billed at twice actual provider token cost. Cached input receives the published cached rate; output and separately reported thinking tokens are billed once. Current built-in tools have no additional customer tool charge. Prices and effective dates are published on Pricing.

Atomic reservations prevent overspending. Definite pre-generation provider rejection releases a reservation. Ambiguous usage retains reserved funds for reconciliation instead of inventing a charge or refund. Unused settled reservations are released. Refunds revoke only available credits and flag unrecovered amounts for review without making a wallet negative.

Captured payments grant access and credits after server verification. Check a pending payment before another attempt. Financial records are retained for receipts and reconciliation. Bank conversion, taxes and provider availability may vary.

- [Current pricing](https://warhoundai.com/pricing/)
- [Website terms](https://warhoundai.com/terms/)

## Retained benchmark run details

The published source-code test used 50 flawed examples and ten fixed controls on one sandbox origin, grouped into 6 source bundles. Backend examples were inert text, not executable public services. The run used Plus Soft with Gemini 3.8 Flash. Ground-truth labels were withheld from the model.

Scoring compared the model’s root-cause explanation and source evidence with the planted issue. A different warning alone did not count. The corpus and scoring used AI-assisted WarHound review, not an independent assessor. Every scored case uses its first complete verdict.

There was 1 incomplete larger preflight request; its exact finish reason was not retained. Batch sizes were reduced, finding output bounded, Soft reasoning lowered and terminal handling for output-limit errors added. The final batch reached the normal five-hour quota after saving candidates. Recovery returned those candidates without more inference. One audited internal allowance refresh funded a complete rerun; weekly usage and historical records were retained. No accuracy-based best-answer selection was used.

Completed batch times ranged from 38.79 to 94.43 seconds, with a 61.01-second median and a 63.37-second arithmetic mean. Completed batches used 139,071 weighted tokens; all attempts used 172,169. This was not a Plus throughput test or a matched competitor speed test. Full mode, active exploitation, whole-site coverage and automatic-fix quality were not evaluated.

- [All case results and source hashes](https://warhoundai.com/benchmarks/source-review-50.json)
- [Interactive benchmark report](https://warhoundai.com/mini-hound/benchmarks/)

## Availability, support and assurances

No independent security certification, public bug-bounty reward, service-level agreement or complete language-coverage matrix is announced. Customer counts, revenue and startup-program approvals are not implied by the website. Contact founder@warhoundai.com for support and security reports. Keep reports private and avoid collecting other people’s information or running disruptive tests.

Product illustrations and sample conversations show the workflow; they are examples rather than findings from a visitor’s website. Planned BloodHound workflows and integrations are labeled separately from the live Mini Hound release.

## Credits, safety and recovery

Mini Hound has no weekly, five-hour, fixed tool-call, iteration or total scan-duration quota. Productive investigation can continue while credits, permission and provider availability remain. Reasoning controls investigation depth and never grants additional tool permissions.

Finite worker invocations and browser sessions checkpoint ongoing work for continuation. Source and context memory bounds, concurrency controls, no-progress detection, scope checks and provider restrictions protect infrastructure. Original evidence remains distinct from model summaries.

Completed paid responses can be recovered without another inference charge. Further model calls consume credits. Exhausted wallets pause saved investigations for resumption; cancellation and revoked authorization stop further dispatch.

- [Current plans](https://warhoundai.com/pricing/)

## BloodHound roadmap and internal research

BloodHound is the larger product in development. Its diagrams and conceptual API examples describe planned workflows, not a released API specification or a supported-coverage matrix. Public BloodHound assessments, continuous monitoring, automatic production changes, editor plugins and browser extensions have no announced release date.

Administrator-only collection, review and research roles are internal prototypes. Their existence does not establish public multi-model agent access, active exploitation or browser validation. Broader product coverage, compatibility and release configuration will be documented when available.

BloodHound is an agent product, not a newly released foundation model. No proprietary model weights or training methods are claimed.

The intended workflow keeps permission, evidence, uncertainty and human decisions distinct. Suggested changes need testing against the application’s roles, transactions and deployment. A polished interface or agreement between models does not independently prove security.

- [Product status](https://warhoundai.com/product-status/)

## Performance targets

Our product targets are 99% accuracy, 5× faster reviews, 99.9% uptime and 4× lower cost. These are development goals, separate from measured results. They are not achieved performance figures, an uptime guarantee or a service-level agreement.

The homepage labels this group Performance targets. Current measured source-code results remain 49 of 50 seeded issues detected and zero of ten fixed controls flagged. We do not relabel that result as 99% accuracy, 49/49 or 100% detection.

How each target will be evaluated

| Target | Definition and evidence |
| --- | --- |
| 99% accuracy | A future accuracy evaluation must define correct and incorrect decisions, positive and negative cases, and its complete test set. The current 98% seeded-issue detection rate is a different measure. |
| 5× faster | The goal is one-fifth the review time for a named comparable workload, baseline and quality threshold. No matched speed comparison has established this result. |
| 99.9% uptime | A service-availability goal, evaluated against a named endpoint, monitoring window, check frequency and downtime rule. No observed 99.9% WarHound uptime result is published. |
| 4× cheaper | The goal is one-quarter of the total cost, or 75% less, for comparable work and billing units. A baseline and usage assumptions must be published before this becomes a measured cost claim. |

## Detection results

A detection rate counts matched, predefined issues divided by every seeded issue in that test. Missed and inconclusive cases remain in the denominator. We retain the full case list and scoring method.

The current published run found 49 of 50 seeded source-code issues: 98% detection. None of ten fixed controls were flagged as security issues. These are separate measures. Source-code detection is not a claim about every vulnerability or every website.

- [See every case](https://warhoundai.com/mini-hound/benchmarks/)
- [Download the test results](https://warhoundai.com/benchmarks/source-review-50.json)

## Timings and workloads

Timing claims identify the task, model, mode, sample size and statistic. We distinguish an average from a median and a batch from an individual review. The retained run has a 61.01-second median per batch of ten source examples.

The cited CoddyKit article reports 5–15 minutes for a Strix changed-file PR scan, 2–4 hours for a medium-sized app with 50–100 endpoints, and 6–12 hours for a large enterprise app. This is a third-party description, not a timed run conducted by WarHound AI.

These tasks use different scope and coverage. We do not turn separate source-review and pentesting measurements into a head-to-head speed multiplier.

- [Read the cited Strix timing article](https://www.coddykit.com/pages/blog-detail?id=512885&slug=strix-the-ai-penetration-testing-tool-that-hacks-your-app-before-real-hackers-do)

## Prices and included usage

Our comparison uses published starting prices in USD, with the billing unit and extra usage fees alongside the price. On 3 November 2026, Mini Hound Plus costs $4.99 for one month with manual renewal. Strix lists Pro at $29 per seat/month and says pentests are billed separately, pay per test.

This comparison covers the named paid plans. Strix also offers an open-source edition; self-hosted costs are outside this price comparison. A lower starting price does not mean identical features or coverage. Mini Hound currently reviews public source; Strix advertises API and web app pentesting and PR reviews. Provider prices, taxes, currency conversion and available plans can change.

- [Current Mini Hound plans](https://warhoundai.com/pricing/)
- [Strix pricing source](https://www.strix.ai/pricing)

## Uptime

An uptime result needs a named service or endpoint, monitoring source, time window, check frequency and a definition of downtime. A target or a hosting provider’s service commitment is kept separate from observed WarHound availability.

We publish an uptime percentage once its monitoring record is available. The source-code benchmark does not measure service uptime.

## Updates and corrections

New runs are published as new versions with their complete results. A successful subset does not replace a larger test or remove its misses. We retain dated historical results and label the relevant model and method.

Send a correction or a reproducible report to founder@warhoundai.com. We update claims when the supporting evidence changes.

- [Contact WarHound AI](https://warhoundai.com/contact/)