Pentest and Security Testing · Deep Hat
Autonomous penetration testing: authorized, continuous, with a human at every gate
AI agents plan and run authorized tests against your own systems, continuously. Deep Hat, Kindo’s cybersecurity model, does the reasoning inside a governed agent harness. Scope, permissions, and every action: controlled and recorded. Anything disruptive waits for a person.
AUTHORIZED SCOPE
TESTER-SCOPED PERMISSIONS
HUMAN-IN-THE-LOOP GATES
FULL RECORD
AIR-GAPPED SUPPORTED
Automatic · in scope
Recon and enumeration
Read-only mapping of the authorized target
Gate · a person approves
Exploit or move laterally?
The agent stops and asks. Declined steps never run.
Approved
Step runs in the harness
With the tester’s permissions, no broader
Every step
Logged against the agent, the tester, and the Group
Command, result, approval, reasoning. In Kindo’s Audit Trail.
Runs automatically
Waits for approval (set per Group by an admin)
Which actions gate is an admin setting. Read-only enumeration runs on its own; exploitation and lateral movement wait for a human.
What is autonomous penetration testing?
Autonomous penetration testing is penetration testing by AI agents that reason about a target, choose techniques, execute them, and report findings, within a scope and rules of engagement the organization sets. On Kindo it runs on Deep Hat, Kindo’s cybersecurity model, as governed agents in the agent harness.
Not the same as
Vulnerability scanning
Known signatures. No reasoning, no chained techniques.
Not the same as
Manual pentesting
Periodic. Limited by tester hours. Months between engagements.
This page
Autonomous, governed testing
Continuous. Agent-driven within scope. A person approves anything that could cause harm.
Done well: continuous, repeatable testing between engagements. Done badly: an unsupervised agent with offensive tooling and production access. The difference is governance.
Why does AI penetration testing need a control layer?
An agent capable of attack-path analysis and exploitation must be constrained more tightly than any other agent in the enterprise. Five things must be true before it runs.
Explicit, enforced scope
Targets and rules of engagement set by an admin, not a prompt.
Tester-scoped permissions
No broader than the authorized tester’s. From your identity provider.
Approval gates
Before any step that could alter, disrupt, or exfiltrate.
Governed execution
Inside Kindo’s agent harness. Gateways control data and model access.
A complete record
Every command, result, approval. Reviewable after the fact.
A hosted general-purpose model
What security teams need
Not trained for adversarial reasoning
A model built for adversarial reasoning, attack-path analysis, and vulnerability research
Often refuses authorized offensive work
Built for authorized security work, with a person approving disruptive steps
Sends your target details to a third party’s cloud
Runs where you control it: Kindo cloud, your cloud, on premises, air-gapped
No scope, no gates, no record beyond a chat history
A harness that enforces scope and gates and keeps the record
What problems does Kindo solve?
Four gaps that annual engagements and scanners leave open. A governed agent closes each; scanners and external testers stay in the picture.
Gap
Months between engagements
Changes ship. CVEs drop. Nothing re-tested until next year.
Kindo → Scheduled re-tests after changes, deployments, disclosures. Same scope and gates every run.
External engagements continue, from better ground.
Gap
Tester hours spent on recon
Enumeration eats the budget before the interesting work starts.
Kindo → Deep Hat maps the target within scope and proposes a plan with its reasoning.
Your red team reviews and directs.
Gap
Detections never validated
Rules tuned against assumptions, not realistic techniques.
Kindo → Gated, in-scope techniques give the SOC real signals to validate against.
Your SIEM and EDR keep detecting.
Gap
Findings without evidence
Lists, no reproduction steps, no path to a ticket.
Kindo → Ranked findings with evidence and reproduction steps. Tickets to ServiceNow or Jira.
Remediation can run as a Kindo agent under the same governance.
How does Kindo run pentest automation?
Deep Hat reasons. The agent harness executes. Kindo’s three core capabilities control what the agent can see and do. Five stages, every one recorded.
Scope and authorization
Admin sets targets, rules of engagement, and the Group allowed to test. RBAC from Okta, Entra, or LDAP caps the agent at the tester’s permissions.
Planning and recon
Agent maps the target within scope via the AI Data Gateway (CMDB, cloud accounts, scanners). Proposes a plan with its reasoning.
Human-in-the-loop gates
Before anything that could alter, disrupt, or exfiltrate: stop and ask. Which actions gate is set per Group.
Execution and record
Approved steps run in the harness. Every command, result, and approval logged against agent, tester, and Group.
Report and remediate
Ranked findings, evidence, reproduction steps. Tickets to ServiceNow or Jira. Optional governed remediation.
The model
Deep Hat, Kindo’s cybersecurity model
Built for red teams and security operators doing authorized work. Runs inside Kindo’s agent harness.
Built for
Adversarial reasoning, attack-path analysis, vulnerability research
Runs in
Kindo cloud, your cloud, on premises, air-gapped
Controlled by
RBAC, AI Data Gateway, AI Model Gateway
Leaves your environment
Nothing. Tooling and findings stay where Kindo runs.
Automatic
Read-only enumeration and mapping, within scope
Waits for a person
Exploitation, lateral movement, anything that alters, disrupts, or exfiltrates
Who sets gates
An admin, per Group. Policy, not a prompt the agent can argue with.
Who runs tests
Members of the authorized Group, with their own identity-provider permissions
Recorded
Every command, result, approval, and the agent’s reasoning. Audit Trail.
Results go to
A ranked report; tickets to ServiceNow or Jira
What will you gain?
Continuous pentesting between engagements: same scope, gates, and record every run. Results that trend.
Coverage
No gap between engagements
Scheduled runs cover the months an annual test leaves open.
Timing
Re-test on change
Deployments, config changes, new CVE disclosures.
Output
Findings you can act on
Ranked. Evidence and reproduction steps. Routed to tickets.
Trend
Results over time
Same scope and record each run, so posture is comparable.
Before
One engagement a year
Hours on recon. Findings as a list. Months with nothing re-tested.
With Kindo
Continuous, governed testing
In-scope tests on a schedule and on change. A person approves every disruptive step. Every action on record.
External engagement
Starts from better ground
Testers begin with Kindo’s findings and trend data, not a blank map.
Who is autonomous penetration testing for?
Red teams, blue teams and SOCs, and application security teams that need testing more often than engagements allow, and need it governed.
Red teams
Scale recon and attack-path analysis
Deep Hat maps and proposes. The team directs and approves.
Blue teams and SOCs
Validate detections against real techniques
Pairs with SOC Automation on the same platform.
Application security
Continuous testing of internet-facing systems
Re-test on every deployment, within an authorized scope.
Systems you own or are authorized in writing to test
A need to test between engagements, not only annually
Offensive tooling and findings must stay in your environment
Why choose Kindo for AI red team and security testing?
A model built for adversarial work, and a harness that governs it. Comparison by category; your scanners and external testers keep their jobs.
Approach
Trained for adversarial work
Where it runs
Scope and permissions
Gates
Record
Hosted general-purpose model
No; often refuses
Third-party cloud
Whatever the prompt says
None
Chat history
Point pentest tool
Varies
Vendor cloud or appliance
Tool-specific
Varies
Tool log, separate from your other AI
Manual engagement
Yes (people)
On site or remote
Contract and rules of engagement
Tester judgment
Report
Kindo with Deep Hat
Yes
Kindo cloud, yours, on premises, air-gapped
Admin-defined scope; tester’s RBAC permissions
Per-Group approval before disruptive steps
Every command, result, approval, reasoning. Audit Trail.
The second difference is where the test lands: in the same record as everything else Kindo knows about the agent.
From the pentest agent
A test step
approver, tester, Group, reasoning
One record in Kindo
Read by
One history, not a tool’s log somewhere else
Kindo cloud · your cloud · on premises · air-gapped
That record is what makes this governed testing, not only automated testing. Gartner lists agent guardrail management and agent security component verification among the mandatory features of an AI agent management platform, and names automated security testing and adversarial prompting as mandatory features of AI application security.
Kindo’s view: governed pentest automation maps to the former and supports the testing portion of the latter.
Sources: Gartner, Market Overview: AI Agent Management Platforms, July 2026; Gartner, Emerging Market Quadrant for AI Application Security, September 2026.
How do you get started?
Request a demo of Deep Hat running an authorized test against a scope like yours. The plan it proposes. A gate stopping a disruptive step. The record every action leaves behind.
Runs on the Kindo platform with Deep Hat. Kindo’s pre-built pentest agents included. Kindo cloud, your cloud, on premises, or air-gapped.
Go deeper on pentest and security testing
Featured · Webinar
Webinar - Deep Hat v2: A Panel on the Future of Cybersecurity AI
Panel with Kindo and Aireon experts on the Deep Hat V2 model for threat hunting and security automation.
View resource →
Product
Deep Hat, Kindo’s cybersecurity model
What it is built for and where it runs.
Use case
SOC Automation
Triage, enrichment, and response on the same platform.
Use case
Air-Gapped and Sovereign AI
Run Deep Hat and Kindo with no external calls.
