LMRLINE MEDIA RESEARCHINDEPENDENT R&D / JAPAN

AUTONOMOUS PENETRATION RESEARCH

AI Self-Growing Penetration System

Experience from testing,
into your next defense.

It gathers, it tests, it learns from the results.
Each finding becomes knowledge for the next test.

A local GLM, labeled labs and a growing RAG database.
LMR is building a foundation where penetration testing keeps improving.

PRIVATE COMPUTE. VERIFIABLE GROWTH.Implementation and test records, per the spec ↓
RESEARCH SNAPSHOT2026.11.17 / values from Detailed Specification v1.0
30Cumulative cycles
6,463Experiments
20Lab environments
36Labeled matchers

In-specification tallies from our labs and in-house environments. Not a third-party evaluation, and not a performance guarantee for customer environments.

01 / BEYOND A FIXED SCENARIO01

Repeat the tests.
Update the knowledge too.

Rather than only running fixed scenarios, the GLM chooses the next test from what it observes. Results are scored and fed back into the next decision.

“Self-growth” here does not mean updating model weights. It is a mechanism that updates the knowledge, tools and prompts in continuous use, from test results.

01

MEMORY

Grow the knowledge

Conditions, findings and reasons for failure accumulate in the RAG database. Last time’s lessons shape the next test plan.

↗
02

TOOL

Grow the testing tools

New detection patterns go into the matchers. Checking against known answers widens what can be verified.

↗
03

PROMPT

Grow the judgment

Scoring feedback feeds the standing instructions, guiding decisions to adopt, retry or stop.

↗

An update does not guarantee improvement every time. Evaluation and demotion separate real gains from mistakes.

02 / THE AUTONOMOUS CYCLE02

From collection
to the next test.

Information is collected from seven sources every ten minutes. A daily test cycle tries the knowledge, evaluates it and makes it usable.

Select a step to see the role of each stage.

Mechanism demo / not connected to the live system

COLLECT / 10 MINUTES

Take in new techniques, broadly.

Public reports and papers, product vulnerability advisories, reproduction tests, expert knowledge, AI attack intelligence, autonomous discovery and our own observations — seven sources, collected continuously.

OUTPUT

Techniques and their test status saved to the collection DB

Test results flow into the next cycle ↗
Every 10 min collect new informationDaily reset the labs and cycleOn completion the resident R3 scores
03 / SEPARATE TO VERIFY03

The one that tests, and the one that judges — apart.

What the AI finds is not taken as truth as-is. Roles and scoring criteria are kept separate to confirm the result.

R1

EXPLORE

Attack & explore

From past knowledge and observations, the GLM selects the next test. It uses seven tools and records what it finds.

Observe → Reason → Act
R2

REVIEW

Review candidates

Candidate techniques are reviewed for quality. R2 decides to adopt, retry or stop, carrying the cycle on.

Candidate → Review → Decide
R3

SCORE

Scoring against answers

A resident process detects completed experiments, compares against 36 matchers and the labs’ ground truth, and records false positives and misses too.

Result → Compare to truth → Evaluate
04 / KNOW WHAT IS TRUE04

Finding it is
not the end.

The same technique is compared across a vulnerable and a fixed environment. We evaluate both finding what is there and not flagging what is not.

WEAK VARIANTVulnerable

Can it detect a real issue?

Technique→Issue present→TP / miss

Detecting an issue that is in the ground truth is a true positive. Issues it misses are recorded too, for improvement.

FIXED VARIANTFixed

Does it avoid reporting a non-issue?

Same technique→Fixed→Correct negation / FP

Techniques that still fire on a fixed target are caught. Knowledge that causes false positives is demoted and excluded from real-world use.

20 LABS / HYBRID ENVIRONMENTS

Web apps, REST APIs, authentication and authorization, file handling, mail servers and more.
Docker and VM labs are used as needed, testing where the answer is known.

MEASURED IN THE LAB

Keep the hits and the misses.

Spec tally / 2026.11.17
99.5%Detection precision

TP ÷ (TP + FP)
4,175 ÷ 4,194 ≒ 99.5%

True positive TP4,175
False positive FP19
Miss miss965
Correct negation correct negation1,271
Inconclusive inconclusive22

6,452 scored in total, 11 fewer than the number of experiments. The 99.5% is the precision of what was detected, not the share of all issues found. For reference, TP ÷ (TP + miss) is about 81.2%. All are values calculated from the specification’s lab tallies.

05 / KEEP THE KNOWLEDGE CLEAN05

So that wrong knowledge
is not grown.

More knowledge is not automatically good growth. Reverse-testing on fixed environments, and a gate before anything enters the knowledge pool, hold back the return of errors.

01

Review the knowledge you hold.

Existing reproduced techniques are re-tested on fixed environments. Any false positive is demoted and dropped from the candidates for real-world use.

RECHECK / DEMOTE
02

Stop errors from coming back.

Re-registering knowledge from past successes is checked as well. Problematic records are excluded and the knowledge is rebuilt from the sound ones.

CHECK BEFORE PROMOTION
E1Reported / collected
→
E3Reproduced in-house
→
E4Customer-environment testing stage

E1–E4 are this system’s own status categories. E4 is the definition of the final stage; it does not indicate a track record of customer engagements.

06 / PRIVATE AI ARCHITECTURE06

Compute and knowledge, kept in-house.

The GLM’s judgment, the labs’ observations and the RAG’s memory in one cycle. Each stage is linked by records and state.

COLLECT

Seven-source collection

Public sources / autonomous discovery / own observation

→
STORE

Collection & test-state DB

PostgreSQL

→
OBSERVE

Progress & results view

API / streaming

LOCAL INFERENCE

GLM-5.2

B300B300B300B300

An inference base on four GPUs

AUTONOMOUS LOOP
R1 exploreR2 reviewR3 score

Oversees lab testing, evaluation and knowledge updates

RAG / QDRANT

Shared knowledge pool

Verified techniques and past know-how,
managed by state for search and reuse

VERIFYIsolated labs

weak / fixed · Docker / VM

Test in labs → gate → permitted application
APPLYScope-limited real-world testing

Scope check / run log / report

A logical diagram based on specification v1.0. It does not represent wiring or the physical placement of machines. Current real-world testing is limited to two LMR-managed domains.

07 / FROM FINDINGS TO ACTION07

What happened,
with the evidence.

Live progress and commentary, the findings and a PDF report, all connected. The course of the assessment becomes material for planning your response.

EVIDENCE → FINDINGS → REPORT

Watch it run,
take the results with you.

  • Streams test progress and commentary
  • Organizes findings and run logs
  • Generates a PDF report
  • Shares to Slack for response planning

The panel at right illustrates the kind of output. It is not an actual customer report.

LMR ASSESSMENT REPORT
REPORT PREVIEW / SAMPLE

Assessment Report

Scope, conditions and findings


01Test overview
02Observed issues
03Run logs and evidence
A test record that informs your next decision.
CONTROLLED EXECUTION

Scope limited / destructive actions refused / DoS refused / abnormally long runs watched

Designed for testing our own systems, or environments whose authorization and scope have been confirmed.
08 / QUESTIONS

Good to know.

Does “self-growth” mean retraining the LLM’s weights?+

In this system, the RAG-DB knowledge, tools such as the matchers, and the prompts in continuous use are updated from scoring results. It does not mean retraining the model’s own weights.

Does it remove the need for human review entirely?+

In the specification, collection, testing, scoring and knowledge updates run as an automatic cycle, while people watch and handle anomalies. Setting the scope at adoption, confirming authorization, key decisions and operational management are still needed.

What exactly is the 99.5%?+

It is the detection precision computed from the 4,175 true positives and 19 false positives in the specification. It is not a detection rate, which would include the 965 misses. It is a tally under lab conditions and does not guarantee the same figure in a customer environment.

Can it be deployed in our environment?+

We review the target systems, test scope, labs needed, compute and reporting/operations to design a configuration. Under the current specification, real-world testing is limited to LMR-managed domains; extending to other environments requires individual design and confirmation.

Are the figures shown here real-time?+

The figures on this page are tallies recorded in “AI Self-Growing Penetration System, Detailed Specification v1.0”, reviewed on 17 November 2026. The cycle display is a demo explaining the mechanism, not a live console of the running system.

LET’S BUILD THE NEXT CYCLE

From a one-off assessment
to a test base that keeps growing.

Tell us the system you want tested and the challenges you face.
Together we will consider how to connect LMR’s research base to your environment.

What to include: target system / what you want tested / number of users / existing equipment / desired operation