The definition
Agentic security testing is security testing performed by a set of autonomous software agents, each specialized in one slice of the attack surface, coordinated by an orchestrator and gated by a human authorization. It has four defining properties: it is continuous rather than time-boxed; it relies on non-destructive detection as its default evidence; every test carries an explicit safety level that decides where it may run and who approves it; and each finding is delivered with a remediation and closed only by re-test.
The agents do what a good security team does — reconnaissance, testing, triage, remediation, verification — but continuously, in parallel, and with a written record of every decision. What makes it agentic is not the use of AI in a scanner; it is that the units of work are agents with a scope, a safety level and an approval path, and that an orchestrator, not a schedule, decides what runs.
What it is not
- Not an automated pentest. A pentest proves impact by exploitation; agentic testing establishes exposure without exploiting, and escalates to exploitation only on staging, under signature.
- Not a vulnerability scanner with a chatbot. A scanner has no scope gate, no safety levels and no remediation; it is on or off.
- Not autonomous offense. Agents never choose their own targets or intensity. Scope comes from a signed document, intensity from the S0–S3 scale, and the person accountable for production applies the fixes.
At SWR7 the swarm is made of specialized agents under one orchestrator. Each agent owns a domain: TLS and certificates, HTTP security headers, authentication flows, injection surfaces, cross-site scripting, dependencies and known CVEs, exposed files and paths, subdomains, leaked credentials, traffic anomalies. The orchestrator decides what runs, in which order, and — crucially — what is not allowed to run.
How it differs from a penetration test
A classic pentest is a time-boxed engagement: a team spends days or weeks on your systems, exploits what it can, and hands you a report. It is valuable and it is not going away. But it has three structural limits:
- It is a snapshot. The day after the report, a deploy can reopen everything.
- It proves by breaking. Exploitation is the evidence, which is why pentests run on staging or in narrow windows.
- It ends with a list, not a fix. Remediation is your problem, often in a hundred-page PDF.
Agentic testing flips those three points. It runs continuously rather than once. It relies on non-destructive detection — evidence that a flaw exists without exploiting it — so it can safely look at production. And for each finding it produces a ready-to-apply remediation addressed at the root cause, not the symptom.
How it differs from a vulnerability scanner
Scanners are fast and cheap, and they are a component of any serious program. What they lack is judgment: they cannot decide that a finding is a false positive in your context, they cannot chain two low-severity issues into one critical one, and they cannot write the fix. Agents can, because each agent reasons about its slice and the orchestrator reasons about the whole.
The other difference is safety levels. A scanner is either on or off. In the SWR7 swarm every test carries a level from S0 (passive, always allowed) to S3 (intrusive, staging only, signature required) — the full definition is on the safety levels page. The default is S0–S1: find without breaking. Anything above is gated behind explicit approval.
Where the human stays
"Agentic" does not mean unattended. Three gates keep a person in charge:
- Scope. Nothing is tested outside the targets you own or are authorized in writing to test. The allow-list is enforced by the orchestrator, not by good intentions.
- Authorization. No test starts before the Rules of Engagement are signed. Intrusive levels require an additional, explicit go.
- Remediation. The swarm proposes; your team applies. On your systems SWR7 is propose-only.
What you get
A continuous view of your exposure, a PDF report for each run with the fix for every finding, alerts within minutes when something changes (an expired certificate, a new subdomain, a header that disappeared after a deploy, an unusual traffic peak), and a client area where the status of each protected site is visible at any time.
Frequently asked questions
Is agentic security testing a replacement for a penetration test?
Not necessarily. Many regulated environments still require a periodic human-led pentest, and SWR7 runs intrusive levels on staging when you ask for them. Agentic testing covers the gap between two pentests: the other eleven months, in which your surface changes and nobody is looking.
Does it run on production?
Yes, at levels S0–S1 (passive and non-destructive). Controlled tests (S2) run on staging — on production only in a window you authorize in writing. Intrusive tests (S3) run on staging only.
Does SWR7 apply the fixes for us?
No — your team does. SWR7 generates the remediation — the concrete change, on the root cause — and verifies afterwards that the finding is gone. On client systems the swarm is propose-only.