Offensive pentest with AI — findings you can trust.

Security testing cannot keep up — and a whole new attack surface has appeared.

Three simultaneous pressures make the classic pentest model stop scaling.

Regulatory context. NIS2, DORA, the CRA (Cyber Resilience Act) and the EU AI Act turn security testing from an option into a periodic, auditable obligation. Whoever offers verified and traceable results, on-premises when needed, wins the regulated market.

One platform, two components, two delivery models.

A control plane (API, multi-tenant, billing, reports) and a scanning-and-verification engine. Both delivered as SaaS on scanope.com or as an on-premises stack, from the same set of images.

The 11 audit types — dispatched by the engine

No “silent fallback”. Each type has its own validator that checks the target's shape (URL / IP-CIDR / git repo / OCI reference) before dispatch — an unknown type is cleanly rejected, it does not accidentally become a site-audit. More than 50 scanning modules, mostly proprietary, with very few external dependencies.

Findings you can trust — verified, not just detected.

This is where the real value sits. Every vulnerability goes through an independent verification process before it reaches the report, and the result comes out with a clear verdict, a confidence level and a false-positive probability — and prioritised by real exploitability. The report contains what is real, not a dump of unverified alerts.

flowchart LR
  A["Raw finding"] --> B["Independent verification"]
  B --> C{"Verdict"}
  C -->|"confirmed"| D["Report: verdict + confidence<br/>+ false-positive probability"]
  C -->|"plausible"| D
  C -->|"refuted"| E["Dropped from the report"]
  classDef n fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef good fill:#dcfce7,stroke:#16a34a,color:#14532d
  classDef bad fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  classDef dec fill:#fef9c3,stroke:#ca8a04,color:#713f12
  class A,B n
  class C dec
  class D good
  class E bad

Five specialised agents, in a deterministic state machine.

Not an autonomous loop that wanders. A formal state machine that enforces legal phase transitions, with five agents that hand typed results to one another — predictable and auditable.

flowchart LR
  R["① RECON<br/>attack surface"] --> C["② CLASSIFY<br/>classified findings"]
  C --> P["③ PLAN<br/>attack plan"]
  P --> E["④ EXECUTE<br/>scope-gated"]
  E --> V["⑤ VERIFY<br/>verdicts"]
  V --> Rep["⑥ REPORT<br/>final report"]
  classDef s fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class R,C,P,E,V,Rep s

Five gates enforced at runtime, not in documentation.

The offensive loop is driven by models, but every command an agent wants to run passes through hard-stop gates at execution. Control is up-front through policy — scope, tool allow-list, dry-run — plus 2FA at the control plane.

Model sovereignty. The default configuration runs a fully local model (internal endpoint, no external key). The provider changes from configuration only — a local model or an OpenAI-compatible service — with no changes to the platform. The client's data never leaves the infrastructure.

A new category: the security of AI-assistant configurations.

As teams adopt AI assistants, their configuration files become de-facto executable code — and a supply-chain vector. ARTEMIS treats them as such.

From scan to verdict to report — every step is demonstrable.

For regulated buyers, this is the differentiator. Every finding keeps the tool, the exact command, the raw output, the timestamp and the executor as structured evidence — after secrets have been redacted, before storage.

• ✓ Complete evidence — command + raw output + HTTP request/response captured on each finding.
• ✓ Verdict + confidence + false-positive probability — persisted with each finding, exposed by the API.
• ✓ Bilingual RO/EN reports — PDF + DOCX, versioned, with a branding snapshot.
• ✓ Controlled access — hashed download token, expires (14 days by default), download cap.
• ✓ Administrative audit log — who, what action, old/new value, IP, user-agent, timestamp.
• ✓ STIX export — interoperability with external SIEM / threat-intel.

SaaS, on-premises or air-gapped — from the same build.

A four-container stack (edge TLS → control plane → engine → database), packaged two ways. The regulated client runs ARTEMIS in its own infrastructure, fully isolated. A native agent is added that scans from inside the network over a secure connection with a per-agent token.

flowchart TB
  Q{"How sensitive<br/>is the data?"}
  Q -->|"standard"| SAAS["SaaS · scanope.com<br/>platform + AI in CAI cloud"]
  Q -->|"mixed"| HYB["Hybrid<br/>data & models at the client"]
  Q -->|"regulated / isolated"| ONP["On-premises / air-gapped<br/>full stack + local AI"]
  classDef q fill:#fef9c3,stroke:#ca8a04,color:#713f12
  classDef a fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  class Q q
  class SAAS,HYB,ONP a

Frequently asked questions