Home / super-guard
Security hub · super-guard

Fight magic
with magic.

Attackers already use AI. With super-guard, AI defends for you: it picks up the latest vulnerability alerts from CISA and other agencies early, proposes a fix right away, and runs simulated attacks in your verification environment to compare before and after the fix. Once the fix is proven, all that is left is one approval on your phone.

Latest vulnerability alertsInstant fix plans by AIAI-designed attack simulationVerified before vs afterPosture of every systemOne approval for production

AvailableCISA KEV checked every 15 minutes (the other 3 feeds every 6 hours) · AI summaries in 3 languages · LINE / Slack alerts · live production posture cards · per-system, version-aware impact · AI fix plans · automatic before / after comparison in a verification environment

AI does ① to ⑥ automatically; only ⑦ needs a person

~5 minfrom a firewall (NSG) change to the LINE alert
4 / 4patterns detected in a real attack-simulation drill
6 / 6read-only connection checks passed (3 Azure environments)
10,000+vulnerability records; KEV checked every 15 minutes

* Measured and carried out in our own environments, November 2026.

AI closed loop

From alert to verified fix,
AI runs the whole way.

Until now, getting from an alert to "it is fixed and we checked" took a security engineer days. super-guard hands that road to AI. Every step is recorded, and every step can be followed in the console and on LINE.

AI, automatically — up to verification
01📡

Fetch early

Checks CISA KEV every 15 minutes and keeps pulling EU EUVD, US NVD and Japan's JVN.

AvailableCISA KEV every 15 minutes, the other 3 feeds every 6 hours

AvailableThe console shows "published → hub → alerted" times

02🎯

Assess impact

Compares the software and versions of each system, picks out only the vulnerabilities that really affect you, and shows whether they are being exploited.

AvailableMatching by product name, exploitation flag

AvailablePer-system, version-aware impact (the sensor reports the software and versions)

03📝

Fix plan, right away

AI writes a fix for your environment: which version to upgrade to, which setting to change, how to hold the line for now, and how to roll back.

AvailableFor severe vulnerabilities: upgrade, configuration, interim protection and rollback; every plan is checked by code

04🧪

Design the attack

AI picks a simulated attack for this vulnerability from a library of safe probe templates, and sets limits on strength and time.

Available5 generic templates that measure detection

AvailableTemplates that judge from the answer whether an attack worked (.env exposure, path traversal, admin pages, sign-in)

05⚔️

Attack before the fix

First, one attack on the verification environment. Can the vulnerability be exploited? Is it detected?

AvailableRuns on its own in an isolated verification environment; records whether it worked and whether it was detected

06🔁

Fix, then attack again

The fix is applied in the verification environment and the same attack runs once more, comparing before and after.

AvailableFixes what got through, attacks again, and judges before vs after side by side

A person approves — production
07✅

Approve and ship

The verification report arrives on LINE. One approval from you and production gets the same fix, then stays under watch. Important changes are approved with a passkey.

AI does ① to ⑥ on its own. ⑦, the production change, is always approved by a person.

super-guard · verification reportIllustration
CVE-2026-XXXXCaddy path traversal / order systemKEV

Before the fix

Exploited3/3
Detected1/3

Fix

Upgrade Caddy to 2.11.10
+ 1 new detection rule

After the fix

Exploited0/3
Detected3/3
✅ Recommended for productionApprove on LINE →
Illustration: a verification report written by AI (fictional data)
Demo board
The demo board (real screen, fictional data): before the fix 4 of 4 attack types work; after the AI applies the fix and attacks again, 0. Detection does not drop
Why now

Attackers already
have AI.

Attacks got cheap: every company is a target now, not just large ones. They come 24 hours a day, at several systems and clouds at once. Against attackers who use AI, only defense that uses AI keeps up.

🤖

Attacks where AI does 80–90% of the work

In November 2025 Anthropic reported a group, believed to be state-backed, using an AI agent to try to break into about 30 organizations. People made decisions at only 4–6 points per attack. Attacks moved from human speed to machine speed.

Anthropic(2025-11-13)↗

🧬

Malware that rewrites its own code

Google Threat Intelligence (GTIG) found malware that asks an LLM to rewrite itself while running (PROMPTFLUX and others) and malware that has an LLM write its attack commands. Defenses built on fixed signatures work less and less.

Google GTIG(2025-11)↗

BEFORE — defense as usual

  • A separate monitor per system. An attack across environments stays under every single threshold
  • People watch screens. Nobody watches at night or on holidays
  • Hundreds of English vulnerability records a day. Finding the ones that matter takes the whole day
  • "We patched it." Nobody checks whether the attack is really blocked now

AFTER — super-guard

  • One hub links the records of every environment and finds the same attacker across them
  • Rules detect 24 hours a day; only the essentials go to LINE or Slack
  • AI summarizes the exploited vulnerabilities that touch your assets, in 3 languages
  • Simulated attacks before and after the fix prove, in numbers, that it holds
Four values

Loop, show, alert, protect.

🔁

AI runs the loop, from alert to verification

Fetching alerts, assessing impact, writing the fix, attacking before and after, writing the verification report: AI does it in one run. Your part is one approval for the production change.

🛰️

Every environment, and every system's state

Azure, AWS, Google Cloud, Alibaba Cloud, data centers, the office network and business systems on one screen. Cross-environment correlation connects the moves you would miss one system at a time. For each system it also shows whether it is up, reachable and under attack.

📲

In your hand the moment it matters

Alerts go to LINE or Slack by severity, in each person's language; a link opens the details without signing in again (read-only). False alarms such as bot scans are checked against the original records and closed automatically, and people's reasons are kept as "memory".

🛡️

Automation designed to be safe

Simulated attacks go only to your own verification environments with proven ownership, under caps and a kill switch. Blocks in production happen only after a person approves, only inside a reserved band, and roll back when they expire. The hub only signs; the executor in the environment verifies again before it acts.

Per-system posture

Every system:
is it safe right now?

No more staring at a pile of alerts. One card per monitored system: is it usable, is the network through, who is attacking it right now, and which vulnerabilities are still open.

Production

  • Is it usable: external probes plus live access logs (requests, error rate, response time)
  • Is the network through: sensor heartbeats and traffic per line
  • Attacks now: from where, what kind, blocked or let through
  • Vulnerabilities still to fix

Verification environment

  • The same live status as production, plus
  • The attack in progress: each simulated request, how the system answered, and whether it was detected
  • Before-and-after comparison charts
super-guard · posture by systemIllustration
Order systemProduction
Up · sensors 2/2 alive
99.98%availability (24 h)
1,240requests / min
0.2%error rate
180 msresponse time
  • 32 attacks: scans from 3 sources, all blocked
  • 2 open vulnerabilities (1 exploited)
Intranet portalProduction
Warning · responses are slowing down
99.91%availability (24 h)
310requests / min
2.1%error rate
640 msresponse time
  • 5 attacks: login brute force, 4 blocked, 1 under review
  • 0 open vulnerabilities
Order systemVerification
Attack simulation running · CVE-2026-XXXX · run 2, after the fix
Exploited, before → after
3/3 → 0/3
Detected, before → after
1/3 → 3/3
Illustration (fictional data). AvailableProduction posture cards (availability, network, usage, attacks; live). AvailableThe "attack in progress" view of verification environments and the before / after comparison (demo board).
Screens

One console in
Japanese, Chinese and English.

Incidents
Incidents: severity, environments, counts and verdicts, with related memories alongside
Vulnerabilities
Vulnerabilities: four public sources, relevance to your assets, exploitation and AI summaries
Attack simulation (BAS)
Attack simulation (BAS): permits, plans, approvals and results on one screen
Overview
Overview: 24 hours of events, HTTP responses and recent incidents

The console available today. All data on the screens is fictional. Screens shown in Japanese; the console also speaks Chinese and English.

A real drill

When a firewall rule changes
in the middle of the night.

What happened in a real drill on our own AI hub environment on 22 November 2026 (a deny rule for a documentation address, added for two minutes).

01

The hub reads the Azure Activity Log and detects the rule change (S-high)

02

About 5 minutes later the LINE alert arrives, in the person's language

03

The LINE link shows the details on the phone without signing in again

04

A reply acknowledges it. The verdict is recorded in the console, and its reason can be remembered

How it works · for engineers

Outbound only.
Raw logs stay home.

The hub is one independent server per company. Sensors in each environment connect out to the hub (outbound HTTPS with mutual TLS); the hub never signs in to an environment. It receives events, aggregates and configuration only.

🔌

Outbound only

Sensors reach the hub over outbound TCP 8443 with mutual TLS. No inbound port is opened in the environment.

🗄️

Raw logs stay in the environment

Only events, aggregates and configuration go to the hub. No query strings, cookies or bodies; user IDs are pseudonymized in the environment.

👁️

Reading and writing are separate

Monitoring uses read-only identities only (resource-group Reader on Azure, keyless OIDC federation on AWS and GCP). Writing uses another identity, only where you enable it.

🔑

Two keys

The hub only signs a ticket. The executor in the environment verifies it against its own rules before acting. Even a compromised hub can do no more than a time-limited block inside the band.

⚙️

Rules detect; AI summarizes, plans and verifies

22 detection rules and cross-environment correlation decide. Besides summarizing vulnerabilities and proposing analyses, AI writes fix plans and plans the simulated attacks. Scans that ended in a refusal are checked against the records and closed automatically.

⏱️

The hub watches itself

After a stop, the hub reports the downtime and re-checks the records of that time (up to 24 hours). No silent blind spots.

🧪

Simulated attacks hit only your own verification environment

Only verification environments you register and whose ownership is proven (DNS TXT, cloud API) are attacked. Strength and time have hard caps, and a kill switch stops everything at any moment. Only non-destructive probes are used; production is never attacked, only watched.

Connection levels — start read-only, grow step by step

LevelWhat it doesWhat the environment provides
SC1 Assets and configAssets and firewall rules, configuration driftA read-only identity (declared scope only)
SC2 MonitoringEvent collection, cross-environment correlation, sensor livenessA sensor + outbound 8443 + logging turned on
SC3 RehearsalRehearse responses in a verification environmentA verification environment with the same version and config
SC4 ResponseAfter approval, block inside the band and roll back on expiryAn executor identity (writes inside the band only)
Safe AI by design

So that AI itself
is not the way in.

AI goes up to verification; people approve production

AI works on its own only up to the verification environment. Every production change is approved by a person. The AI holds no cloud identity and no signing key.

Attacker text is data

Fields attackers can write (user agents, URLs, login names) are isolated; instruction-like text is treated as evidence of an attack.

Real logs never go to an LLM

The AI gets public vulnerability data and abstract asset facts such as software names and versions, nothing more. Real logs are not sent until an LLM in a Japan region is chosen.

No personal data in memory

A memory with names, contact details or secrets is refused as a whole, never masked. The audit log never holds the text.

A cap on cost

AI use has a daily cap. When it is reached only the AI pauses; detection and alerts continue.

The environment stays in charge

The owner can cut an environment off from the hub at any time by removing the trust or stopping the executor.

What to connect

What to connect, and what to prepare

Every environment starts read-only. Writing (blocks) is enabled only in the environments and ranges you choose. For the AI loop, each system needs a verification environment with the same version and configuration as production (if you have none, we build one with you at onboarding).

EnvironmentRead-only identityMain logsVerification environment
Microsoft AzureReader on the resource groups, for the hub's managed identityActivity Log, VNet flow logs, web access logs, host sign-in logsA separate resource group, same version and config as production
AWSAn IAM role via OIDC federation (SecurityAudit); no access keysCloudTrail, VPC flow logs, ALB / CloudFront access logsA separate account or VPC, same version and config
Google CloudWorkload Identity Federation (viewer roles); no key filesCloud Audit Logs, VPC flow logs, load balancer logsA separate project, same version and config
Alibaba CloudA RAM role + OIDC (ReadOnlyAccess); no AccessKeyActionTrail, VPC flow logs, SLB / WAF access logsA separate VPC, same version and config
Data center / private cloudNone (only the sensor certificate in the management segment)Firewall and proxy syslog, NetFlow, virtualization audit logsVMs or containers of the same version
Office networkNone (an office sensor)UTM and router syslog, DNS, identity logsOptional (a test segment)
Business systems (web)None (the sensor reads the access log)JSON access logs of the reverse proxy (Caddy, nginx and others)Same version and config; if none, built with you at onboarding

The detailed permissions and settings come with the connection guide at onboarding (Japanese and Chinese).

Getting started

Getting started — the hub in a week, live alerts in a month

01

Week 1 · set up the hub: one VM, up in a day

02

Week 2 · connect read-only: SC1 checks the permissions automatically

03

Week 3 · sensors and trial run: SC2; choose recipients and languages

04

Week 4+ · live alerts: AI vulnerability summaries start too

05

Next · the AI loop: verification environment + automatic attack simulation + before/after comparison

Cost guide (per month)

Hub VM (24 hours)about ¥12,000
Logs and flow records per cloudabout ¥2,000–15,000 per environment
AI (summaries, analysis)kept under a daily cap
e.g. Azure 1 + AWS 1 + GCP 1 + officeabout ¥30,000–65,000

Estimates; they depend on log volume and configuration. Verification environments are extra.

Questions

  • AI attacks automatically. Could it hit production? → No. It attacks only verification environments you register and whose ownership is proven. Strength has hard caps, and an emergency stop works at any time
  • The fix is written by AI. Can I trust it? → Each fix is first attacked, applied and attacked again in the verification environment, and the comparison proves it works. Then you approve it
  • What if the hub is taken over? → It has no write access to your clouds and only signs; nothing beyond a time-limited block inside the band is possible
  • We do not want raw logs to leave → They do not. Only events and aggregates go to the hub
  • We have no SOC → The hub watches 24 hours a day and sends only the essentials. It is designed for small IT teams
Availability

Availability (November 2026)

Available

  • Rule-based detection and cross-environment correlation
  • LINE / Slack alerts (3 languages, view links)
  • Vulnerability feeds (KEV checked every 15 minutes; EUVD, NVD, JVN every 6 hours), "published → alerted" times, matching by product name, AI summaries in 3 languages
  • Live per-system posture (external probe, availability, network, usage, attacks)
  • Azure Activity Log and web access logs
  • Attack simulation (BAS, 5 generic templates that measure detection), auto-triage, memory
  • Console in Japanese, Chinese and English

The AI loop

  • Per-system, version-aware impact
  • AI fix plans (upgrade, settings, interim defense, rollback)
  • Attack simulation in a verification environment, before vs after the fix
  • Verification reports and the approval flow to production
  • The "attack in progress" view of verification environments

Set up at onboarding

  • AWS, Google Cloud and Alibaba Cloud connections
  • Sensors for data centers and office networks
  • A verification environment per system
  • Approved response (SC4, band and expiry)
  • Trusted source ASNs and a second contact

Coming

  • Real-log analysis with an LLM in a Japan region
  • Vendor advisories (MSRC, Ubuntu USN and others)
  • Weekly reports delivered automatically
  • As a SaaS

Attackers use AI.
Defend with AI.

Start with one environment, read-only. In a week the hub is in place and you can see where you stand.