Full database read
Public sector portal
A blind SQL injection flaw chained up to full admin rights.
11 findings·61 min
Full breakdown→Product · KLUE
Meet KLUE, your AI security engineer.
One agent that runs continuous pentests across your apps, cloud, code, APIs and CI/CD, and delivers a working proof for every finding. In hours, not days.
The problem
Most tools look for what they already know.
A rule, a signature, a playbook someone wrote years ago. Every scanner shares that engine: it catches what its rules describe, and everything else slips past. There is no signature for a bug your team wrote yesterday, or for two safe steps run in the wrong order. Those are judgments about intent. KLUE reasons about the target instead.
One platform
Six assessments. One graph of findings.
Same reasoning engine, same isolated runtime, same report format. Pick what to test and the agent does the rest.
Autonomous Pentest
Vulnerability discovery, exploitation and chained attack paths, attacker-style.
Vulnerability Assessment
Runtime discovery against live web and API targets.
Static Analysis
Source review for code flaws, hardcoded secrets and dependency risk.
Cloud Audit
AWS, Azure and GCP posture against CIS, SOC 2, ISO 27001 and PCI.
M365 Assessment
Microsoft 365 posture, identity configuration and compliance state.
Threat Intelligence
Unified search across CVE, ATT&CK, malware and credential-leak sources.
Measured, not claimed
We ran the benchmark. KLUE sat above the curve.
Rule-based tools trade coverage against noise. A reasoning engine is not bound by that curve, and it runs fast enough to fit your release cadence.
0%
Precision
Zero false alarms on the public code-review benchmark.
0.5%
Recall
Above the ceiling rule-based tools reach in practice.
0+
CVE-level flaws
Responsibly disclosed in popular open-source software.
~30 min
Per run
Fast enough to run after every release, not once a quarter.
The cadence
Thirty minutes. Not four hours. Not weeks.
How much you catch is half the story. How long it takes is the other half. Most AI testing tools take four to six hours a run, so you can only afford to test once a quarter. A thirty-minute run fits into every release.
| Tool | Recall | Duration | Soft positive rate |
|---|---|---|---|
| KLUEReasoning agent, thirty-minute budget | 82% | ~30 min | ~4% |
| Leading commercial agentCombines several AI models | 75.0% | 4 hr | 6.3% |
| Raw frontier modelA top model on its own | 70.0% | 10 min | 6.7% |
| Open-source agent ASame underlying AI model | 45.0% | 4 hr | 10.0% |
| Open-source agent BSame underlying AI model | 30.0% | 6 hr | 25.0% |
| Open-source agent CSame underlying AI model | 5.0% | 2 hr | 0.0% |
Public benchmark against a well-known target with twenty-two documented vulnerabilities. KLUE figures come from real runs under a thirty-minute budget. Other figures come from publicly reported results.
One
It is how the agent is built, not the AI behind it.
Three tools on the public board use the same underlying AI model and catch 45%, 30%, and 5% of bugs. That gap comes down to how each one is built. KLUE is the difference.
Two
Model choice becomes a cost decision.
On a separate benchmark, KLUE on a cheaper open model matched a top-tier paid model on the most critical bugs, at one-fifth the cost.
Three
Speed is what unlocks it.
Four-hour tests force you to a quarterly schedule. Thirty-minute tests run after every release. That is the bar serious autonomous tools will be measured against.
The landscape
Three categories. The category decides the ceiling.
Every product in this space falls into one of three groups, set by what its engine actually does: matches known patterns, replays known attacks, or reasons about the target. The group sets the cap.
| Capability | KLUE | Vulnerability scanners | Breach simulators | Other AI agents | Human pentesters |
|---|---|---|---|---|---|
| Engine | |||||
| Underlying model | Reasoning | Rules | Simulated playbooks | Mixed agents | Human reasoning |
| Discovers unknown vulnerabilities | – | – | Partial | ||
| Writes custom exploit code | Per target | – | Preset | Preset | |
| Chains findings into attack paths | Multi-step | – | Limited | Some | Manual |
| Delivery | |||||
| Time per engagement | ~30 min to 6 hr | Always-on scan | Hours | Hours | 2 to 6 weeks |
| Output | PDF, data export, working proof | CSV, dashboard | Dashboard | Dashboard, PDF | PDF (weeks later) |
| Free retest after fix | Not applicable | Re-run | Re-run | Extra cost | |
| Operations | |||||
| Runs on every release | Known-flaw scan only | Scheduled | Scheduled | Impossible | |
| Parallel scans | Unlimited | Unlimited, shallow | Scheduled | Scheduled | One per team |
| Vendor model | Proprietary, exclusive | Licensed software | Licensed software | Licensed software | Consulting hours |
Most tools answer “what known weakness might you have?” KLUE answers “what would an attacker actually do here?” The category decides which question can be asked.
Track record
Real engagements. Not pitch decks.
Real targets, real attack paths, every step confirmed with a working proof. Nothing reported on suspicion alone.
Account takeover, no password
Government agency
Any official reached by username. No code, no password required.
7 findings·37 min·CVSS 9.8
Pricing
Plans from $150 per month. Services quoted to fit.
Self-serve KLUE subscriptions for continuous testing, or a scoped quote for pentest, red team and managed security.
Get started
See what an hour findsagainst your own systems.
One hour. Real exploits. Real fixes. No procurement cycle, no consulting engagement. You keep the findings either way.