Skip to main content
AI Pentests puts an autonomous agent to work on your web application. The agent is called Specter. You give it a starting URL and, if you want, written instructions. It explores the application, forms ideas about where it might be weak, tries to prove them, and records only what it can confirm. This is not a scan under a different name. A scan runs a tool you picked, with the options you set, and reports what that tool checks for. A pentest is an assessment: Specter decides what to look at next based on what it just found.
AI Pentests is in Early access. Access is enabled per account, so the feature shows up in your sidebar only if it has been turned on for you. If you don’t see it and want to try it, start with AI Pentests on our pricing page.

​
How a pentest differs from a scan

​
Before you start

Get these in place first, because the agent cannot work around them:
  • A target URL that loads. Preflight blocks the launch if it cannot reach the URL.
  • CAPTCHA and other interactive challenges turned off on the path being tested.
  • Your WAF rules and source IP restrictions adjusted so traffic from 172.236.11.251 gets through.
Use a staging or pre-production target if you have one. Specter runs active checks to confirm what it finds, which means real requests against real endpoints.

​
Start a pentest

1

Setup

Enter the target URL, or pick one of the URL targets already in the workspace. Name the pentest, using a name no other pentest in that workspace has. Add instructions for the agent if you have any.
2

Preflight

We load the target and follow its redirects to check that it answers. If it redirects to a subdomain or to another domain, preflight reports it as unreachable. Go back to setup, enter the URL the target actually lands on, and run preflight again.
3

Review and launch

Check the target the agent will start from, which is the URL after redirects rather than the one you typed. Then confirm the launch acknowledgements, which include confirming that you are authorized to test the target and the systems around it.

​
Guide the agent with instructions

The instructions box is optional and takes up to 2000 characters. Write what you would tell a human tester on their first day:
  • “Focus on the billing flow and account takeover.”
  • “The API behind the app is at https://api.example.com.”
  • “Use the demo tenant. Leave status.example.com alone.”
  • Credentials for a test account, with a note on what that account can do.
Instructions can only be set during setup. Once the pentest is running there is no way to add or change them, so put in everything you know up front.
When you supply instructions, the generated report records that the assessment was guided. A reader then knows the findings came from a directed test rather than an open one.

​
What Specter does during a pentest

Each pentest gets its own environment with a real browser and the tooling a tester would reach for: port and service discovery, injection testing, content and endpoint discovery, and more. The work follows a rough order. Specter fingerprints what the target runs, looks up known issues affecting those versions, builds a list of ideas worth testing, then tries to prove them one at a time. When a line of investigation deserves its own effort, the main agent hands it to a subagent with a specific brief. You can follow each of them in the Activity tab. A pentest runs for up to 48 hours. You can stop it whenever you want, and the findings and activity recorded up to that point stay available.

​
Pentest statuses

​
Review the results

​
Findings

Only confirmed findings reach you. Findings go through a dedicated validator agent before they are stored, and anything it cannot confirm against the live target is dropped and never appears. Severity comes from a CVSS v3.1 vector the agent has to supply, not from a label it picks: Every finding carries evidence, and at least one piece of it has to be proof rather than description. Evidence blocks are one of:
  • a command with its raw output
  • extracted data or a decoded value
  • a screenshot
  • a parameter and its value, as supporting context

​
Activity

Activity is where you watch the work happen. The left side lists the agents, starting with Main and adding subagents as they spawn. Selecting one shows its timeline: what it was reasoning about, each command or browser action it took, the output it got back, and any screenshots. Every tool call comes with a plain sentence explaining why the agent made it.

​
Generate the report

Once a pentest finishes, you can download a PDF pentest report built from the confirmed findings and the recon behind them. It covers:
  • an introduction with background, objective, scope, approach, methodology, and disclaimer, plus the testing window and the source IP used
  • an executive summary with the vulnerability distribution and a list of findings
  • attack path analysis, with the attack vectors used and a flow diagram
  • each finding in full: evidence, description, risk, recommendation, reproduction steps, references, and a classification table with the CVSS score and vector, CWE, and OWASP Top 10 category
  • a remediation priority list, grouped as immediate (within 7 days), short term (within 30 days), and medium term (within 90 days)
  • reconnaissance: open ports and services, technologies identified, attack surface, and observations
Reproduction steps are written so that whoever picks up the report can run them without access to the pentest.

​
Share pentests with your team

Pentests belong to a workspace, so anyone who can see the workspace can see the pentests in it. Starting, renaming, stopping, or deleting one takes Edit access. A teammate with View access can open a pentest and read its findings, activity, and report, but cannot launch one. To let someone run pentests, change their access to the workspace in your team settings. See Teams and roles for what each access level allows.

​
How your pentest data is handled

This section covers AI Pentests specifically. For account level data handling, see the AI data policy. Specter is driven by third party large language models hosted in the US. We do not use your pentest data to train models, and we require the same of our providers. Turning the report into a PDF involves rendering it through a third party service, so the report content passes through that service. We keep a record of each pentest: the target, your instructions, the activity history, the findings, and the report.