Zero Hunt
An on-premise AI security appliance that racks inside your own network. It runs a 10-agent generative penetration-testing swarm, deep-packet AI traffic analysis and continuous mapping against 32 compliance frameworks, with no cloud callbacks and full air-gap support.
What is Zero Hunt?
Zero Hunt is a cybersecurity appliance rather than a service: dedicated GPU hardware plus a pre-configured software stack that you rack in your own datacentre. It ships in two form factors — a desktop unit for smaller organisations and a 2U rack unit for enterprise and datacentre use — and the vendor's promise is that nothing ever leaves your perimeter. There are no cloud callbacks, no telemetry and no external LLM API calls, and full air-gap operation is supported for classified and regulated environments.
What makes the design unusual is that the publisher trains and owns the models it runs. Two in-house families power the engine: ZeroHunt Apex at 27 billion parameters, small enough to run entirely on the appliance, and ZeroHunt Apex Pro at 284 billion for higher-tier hardware and sovereign private-cloud deployments. Both are fine-tuned on a proprietary corpus of roughly 130,000 curated offensive-security examples, reduced from more than 145,000 raw items by near-duplicate removal and decontaminated against the benchmarks the vendor reports on.
Four pillars sit on top. The generative pentest deploys ten specialist agents — recon, exploit, web, credential, post-exploit, pivot, tactic and report — that write attack code for your specific environment instead of replaying a playbook library, backed by 142+ self-evolving skills trained in an internal AI Gym. A separate Generative Zero-Day PoC stage fingerprints the exact software version you run, fetches that version's source, taint-scans it for reachable sinks, then iterates inside a sealed container until a working proof-of-concept fires. Traffic analysis runs a proprietary four-head deep-learning model trained on billions of PCAP sequences directly on the appliance GPU. Compliance maps every finding across 32 frameworks — NIS2, DORA, ISO 27001, GDPR, SOC 2, PCI-DSS, HIPAA, NIST and more — and signs the resulting evidence with ECDSA. A fourth pillar, the AI Remediation Advisor, turns confirmed findings into a prioritised fix plan you can interrogate conversationally.
The publisher is Over Holding Srl, a technology group in Udine, Italy, for which Zero Hunt is one of three ventures.
What it does
- Run autonomous penetration-testing campaigns with a swarm of ten specialist AI agents coordinated by an AI controller
- Have the engine write, run and repair a working proof-of-concept exploit until it demonstrably fires against the target
- Classify network traffic in real time on the appliance GPU, at a stated 2.7+ Gbit/s baseline
- Detect ransomware during the encryption phase, plus covert C2 over HTTPS, DNS tunnelling and domain fronting
- Map every finding continuously against 32 compliance frameworks with weighted gap scoring
- Export ECDSA-signed reports and auditor evidence bundles with a full chain of custody
- Work through a prioritised fix plan with an agentic advisor grounded on the engine's own evidence
When to use Zero Hunt / When not to
A quick filter to help you decide if Zero Hunt is the right fit.
When to use Zero Hunt
- Security teams under a hard data-sovereignty rule, where production telemetry is contractually or legally forbidden from leaving the perimeter
- NIS2 essential and important entities that must produce Title 13 incident evidence on a 24h / 72h / one-month clock
- DORA-regulated financial entities running threat-led penetration testing under Articles 25 and 26
- Organisations already buying annual or quarterly pentests that want continuous validation without adopting a SaaS pentest vendor
- Air-gapped and defence-adjacent environments — prime contractors, classified-data handlers, public administration — where any vendor cloud touchpoint is disqualifying
When not to use Zero Hunt
- Anyone expecting to sign up online: there is no self-service path, no public price and no trial — every purchase runs through an authorised channel partner
- Small teams without a datacentre or a rack, since the product ships as dedicated GPU hardware rather than software you download
- Buyers who need a certified vendor today: the publisher's own Trust Center lists ISO/IEC 27001 as in progress and SOC 2 Type II as planned for 2027
- Organisations unable to carry the paperwork the licence demands — a written authorisation register per target, Rules of Engagement, a scope document and a named accountable security officer
- Developers looking to build on top of it: no public API, no SDK and no technical documentation are published anywhere on the site
How to use Zero Hunt
A typical end-to-end flow, from setup to results.
- Open the contact form on the site and pick a reason — the routes are demo, pricing and licensing, channel partnership, support, press, security disclosure and legal
- Take the 30-minute technical demo, in which the engine is run against a recorded slice of your own stack so you can compare its output with your current tool
- Work through commercial terms with an authorised channel partner, since the product is never sold direct in a covered market
- Sign the Licence Agreement, and the Article 28 Data Processing Agreement where the appliance will process personal data on your behalf
- Before any campaign, assemble the paperwork the licence requires: a written authorisation register per target, Rules of Engagement, an explicit scope document and a named accountable security officer
- Receive the appliance pre-configured and hardened, rack it in the datacentre and connect it to the network
- Place sensors across the network segments you want covered, including OT and ICS networks
- Onboard your estate and let the autonomous campaigns begin — the vendor describes operations starting immediately after connection
- Schedule recurring campaigns by cron, or trigger them on change detection
- Export signed evidence bundles from the Trust Center whenever an auditor or regulator asks
Pros & Cons
Pros
- Genuine sovereignty: the full stack — LLM, embedding model, traffic ML, evidence store and dashboard — runs on the appliance GPU inside your perimeter, with air-gap updates delivered as signed bundles on physical media
- The publisher owns its model weights rather than wrapping a third-party cloud LLM, which removes per-token cost, vendor rate limits and the refusals a general-purpose model gives on offensive work
- Proof rather than opinion: the engine files a runnable proof-of-concept against the campaign, with before-and-after evidence of what it actually did
- Four functions usually bought separately — offensive validation, network detection, compliance and remediation guidance — arrive in a single deployment
- Audit traceability is built in, with ECDSA signature at write time and exports aligned to NIS2 Title 13 and DORA TLPT evidence requirements
- An unusually candid Trust Center that publishes certification status as in progress or planned instead of claiming credentials the company does not hold
- The containment model around the exploit forge is documented in detail — no host access, no internet route, downloaded source sealed as data before the AI reads it, and destruction when the work ends
Cons
- No public pricing at all: the pricing, plans and quote pages simply do not exist on the site, and the real figures sit inside a password-protected commercial kit
- No free trial and no free plan for an end customer — the 90-day loaner appliance is reserved for certified channel partners
- The publisher holds no security certification today: ISO/IEC 27001:2022 is in progress with a stage 1 audit planned for H2 2026, and SOC 2 Type II is only planned for 2027
- Sales are indirect only, so you must find and go through an authorised partner, and the vendor will not sell direct in a covered market
- The entry cost is structurally high because dedicated GPU hardware is part of the deal, with no way to start small and scale up
- Significant preparatory burden on the buyer: authorisation registers, Rules of Engagement, scope documents, a named accountable officer and a minimum 24-month log retention obligation
- A young product against mature competitors — the site itself concedes that Pentera holds the market maturity, the playbook breadth and the partner ecosystem
Pricing & Plans
There is no free plan and no published price. The site carries no pricing page at all — the usual pricing, prices and plans paths do not exist, and none of the 324 URLs in the sitemap points at a tariff. Zero Hunt is sold exclusively through authorised channel partners under a Licence Agreement negotiated per deployment, and the detailed figures live in an authenticated commercial kit that the partner programme describes but does not open to the public. The Terms go further and classify pricing itself as confidential information. Two caveats matter for anyone scanning the site for numbers. The structured data declares a currency of EUR and availability in stock, but carries no amount. And the monetary figures that do appear — such as the 50,000 to 150,000 euros a year quoted for a pentest contractor, or the 50,000 dollars and upward for an annual pentest — describe the buyer's current spend that the product aims to replace, not the price of Zero Hunt. A further figure, 100 million dollars of annual recurring revenue, belongs to the competitor Pentera on a comparison page. To obtain a price you must use the contact form and select the pricing and licensing route.
- No named product plans are published
- a desktop unit positioned for SMBs
- and a 2U rack unit for enterprise and datacentre deployment
- ZeroHunt Apex at 27B parameters for standard on-premise units
- and ZeroHunt Apex Pro at 284B for higher-tier appliances and sovereign private cloud
- Channel tier 1
- Registered — standard reseller margin on registered deals
- no minimum revenue commitment
- Channel tier 2
- Certified — stepped-up margin after one certified technical resource and a first paid deployment
- plus co-marketing budget
- Channel tier 3
- Strategic — top margin for partners with vertical specialisation and committed pipeline
- including joint roadmap access and a named field engineer
Data, GDPR & hosting
A consolidated view of how Zero Hunt handles your data.
GDPR overview
Implementation is concrete and documented. The controller is named as Over Holding Srl, Viale Tricesimo n.200, 33100 Udine, with a designated Data Protection Officer reachable at dpo@zerohunt.ai. The privacy policy cites its Article 6 legal bases, lists the data categories, states retention periods, enumerates the Articles 15-22 rights with a 30-day response commitment, and names the Italian Garante as the supervisory authority. The Trust Center publishes the website subprocessors and their safeguards, and the Terms confirm that an Article 28 Data Processing Agreement is signed as part of the licence whenever the publisher acts as a processor. Because the company is established in Italy, no Article 27 representative is required and none is named. Transfers outside the EEA rely on the EU-US Data Privacy Framework and standard contractual clauses.
Who owns the data?
The Terms split ownership cleanly. Over Holding Srl keeps every intellectual property right in the platform itself — the software, the AI model weights, the AI Gym backtest corpus, the documentation and the Trust Center artefact templates. The customer receives only a non-exclusive, non-transferable, revocable licence, and reverse engineering or extracting the model weights is expressly forbidden. Everything the appliance produces on the customer's own systems belongs to the customer: findings, evidence files, signed reports and remediation artefacts. The publisher takes only a limited licence over anonymised metadata, strictly for EU AI Act post-market monitoring and security incident response, and only where the customer explicitly opts in through the optional sync server — never by default.
Reuse rights
Customers may reuse what the appliance generates without asking permission: the findings, evidence files, signed reports and remediation artefacts produced against their own estate are theirs outright. The Terms set one explicit boundary on that reuse — output must never be presented as a third-party attestation, certification or assurance, because it is the customer's own internal evidence. On the publisher's side, reuse is deliberately narrow. Nothing from a production appliance reaches Over Holding Srl at all, and the company states plainly that it does not see, collect or process data from deployed appliances. The only exception is anonymised metadata for AI Act post-market monitoring and incident response, which requires an explicit opt-in through the optional sync server. Separately, the privacy policy states that no customer data is used for training — the models are trained on public vulnerability databases and synthetic environments.
Data retention & training
Hosting summary
Two perimeters must not be confused. The appliance hosts nothing on the vendor's side: no customer data, telemetry or analytical output leaves the customer perimeter during normal operation, and the publisher states it does not see, collect or process data from production appliances. By design the appliance has zero subprocessors, and air-gap deployment removes even the update channel. The website and sales contact side is separate and disclosed. The Trust Center names three subprocessors: Cloudflare, Inc. for website hosting, CDN and cookieless analytics, under United States jurisdiction with the EU-US Data Privacy Framework and standard contractual clauses; Google LLC and Google Ireland Limited for Google Analytics 4 with anonymised IP and advertising signals disabled, under United States and Irish jurisdiction; and self-hosted mail infrastructure in Italy for correspondence, operated inside the controller's own perimeter with no third-party processor. Website and sales contact data is stated to be stored in EU-hosted infrastructure, and material changes to the subprocessor list are notified with at least 30 days notice.
Things to keep in mind
Risks and trade-offs to weigh before adopting Zero Hunt.
- This is offensive tooling. The licence makes the customer responsible for written target authorisation, Rules of Engagement and a named accountable officer before any campaign, and misuse can expose an organisation to criminal liability under national computer-misuse law
- Autonomy invites complacency. A tool that hunts continuously can quietly become a substitute for security judgement rather than an input to it, and the vendor is explicit that it does not warrant finding all vulnerabilities or satisfying all regulatory requirements
- Output can be mistaken for assurance. The Terms forbid presenting the platform's findings as a third-party attestation, precisely because signed, auditor-ready evidence looks like certification when it is only internal evidence
- Nothing the publisher claims about certification is yet independently verified — the ISO 27001 and SOC 2 lines are declared as in progress and planned, and the AI Act conformity artefacts are only released under NDA
- Export-control and sanctions clauses apply, covering EU dual-use rules, US regimes and the Wassenaar intrusion-software controls, so cross-border deployment is not a purely commercial decision
- The buying path itself limits scrutiny: with no public price, no trial and no named customer, a prospect has little to compare before committing to hardware
- The product is new. The domain was registered in April 2026 and first archived in August 2026, so there is very little public operating history behind the technical claims
Setup & Integrations
Technical difficulty
Physically demanding, operationally light. The appliance arrives pre-configured and hardened, and the vendor describes a three-step path — ship, rack, hunt — with autonomous operations beginning immediately after network connection. What the buyer must supply is rack space or a desktop position, network connectivity, and sensor placement across every segment to be covered, including OT and ICS networks. No development skills are needed: there is no API, SDK or integration work. The real effort is governance — authorisation registers, Rules of Engagement and scope documents must exist before the first campaign. Partners receive a two-day technical certification covering deployment, hardening, sensor placement and evidence export.
Deployment
Integrations
Supported languages
Behind Zero Hunt
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Zero Hunt.
Frequently asked questions
Is Zero Hunt really on-premise, or does it phone home?
How much does it cost?
Is there a free trial or a free plan?
What certifications does the publisher hold?
How do you buy it?
Which compliance frameworks are covered?
Does it work air-gapped?
Is there a public API or developer documentation?
Who is behind Zero Hunt?
What languages is it available in?
Should you pick Zero Hunt?
Zero Hunt answers a narrow question with unusual conviction: what does an autonomous red team look like when nothing at all may leave the building? The answer is a GPU appliance you rack yourself, running models the publisher trained and owns, that generates exploit code for your environment rather than replaying a catalogue, watches your traffic at wire speed, and signs its evidence as it writes it. For a utility under NIS2, a bank facing DORA threat-led testing, or a defence supplier for whom any vendor cloud is disqualifying, that combination has few equivalents.
The engineering story is coherent and specific. Owning the weights is not a slogan here — it is argued through the training corpus, the AI Gym backtesting, the benchmark decontamination and a containment model for the exploit forge that reads like it was written by people who assume the code under test is hostile. The Trust Center is franker than most: it publishes what the company does not yet hold.
That candour is also the caveat. There is no certification today, only an ISO 27001 audit planned and a SOC 2 scoped for 2027. There is no public price, no trial, no named customer, no testimonial and no case study — an unusual absence for enterprise software. The product runs on a domain registered in April 2026 with a first web archive capture in August 2026, so while the publisher dates from 2018, this venture is new. And the only way in is through an authorised partner.
Evaluate it as a serious, early product with a genuine architectural argument, not as a mature platform with a track record you can check.
- Choosing a selection results in a full page refresh.
- Opens in a new window.