Skip to content
Security Review: Possible Rakuten Securities Leak, Up to 15.72 Million Accounts — Production Data in a Test Environment, and Seven Fixes a Small Company Can Apply

Security Review: Possible Rakuten Securities Leak, Up to 15.72 Million Accounts — Production Data in a Test Environment, and Seven Fixes a Small Company Can Apply

Image generated by xAI Grok
15 min read
0
0 comments
553 views

Rakuten Securities copied unmasked production data into a test environment, and data from up to about 15.72 million accounts may have leaked. This article lays out the full timeline, four root causes shared by eight incidents in the same month, and concrete fixes for test data, access auditing, vendors, AI code scanning and Japan's personal information law.

I am a Rakuten Securities customer too, and I got the apology email this week. On November 17, 2026, Rakuten Securities announced that a third party had gained unauthorized access to its test environment. Information from up to about 15.72 million accounts may have leaked. For up to about 4.47 million of those accounts, the data includes name, date of birth, address, employer, withdrawal bank account, and self-reported annual income and financial assets (Rakuten Securities notice, Impress Watch).

As of November 17, Rakuten Securities had found no leak of passwords or PINs, and no fraudulent trades caused by the incident.

My judgment: this one is serious because of the fields, not the headcount. Income, assets and a withdrawal account in one record are enough to run targeted fraud. The 227,216 records Sumitomo Mitsui Banking Corporation (SMBC) disclosed the same month were email addresses only, a different level of risk (SMBC notice).

This article has three parts: what happened; what I saw with test data more than ten years ago, how I judged it then, and how I see it now; and technical fixes a small company can apply as written.

Timeline

The data may have left on November 11 and was detected five days later. The announcement came the day after detection.

Date (2026)Event
Nov 20, morningMore than 15,000 people had registered on a site preparing a class action; no suit filed yet (FNN)
Nov 19Japan's Financial Services Agency (FSA) issued an order to report, under Article 56-2(1) of the Financial Instruments and Exchange Act and Article 146(1) of the Act on the Protection of Personal Information (Business+IT)
Nov 18Rakuten Securities told reporters that a system upgrade required checks against production data, so part of its customer data was copied to a test server without masking, through a formal internal approval (ITmedia); a lawyer opened a site recruiting class-action participants the same day
Nov 17Announced a possible leak of up to about 15.72 million accounts: about 4.47 million with full personal data, about 11.25 million with customer codes and login IDs; former customers included
Nov 16Detected the unauthorized access, cut it off and applied fixes
Nov 11Part of the information managed in the test environment may have left

The FISC security guidelines used by Japan's financial industry do not recommend testing with real data (Tokyo Shimbun).

The test data I have seen

Rakuten Securities' test environment held unmasked production data. I saw this practice more than ten years ago, and it was common.

What happened then. More than ten years ago I was doing technical work in Shanghai. Many companies ran their tests mostly on real data. Companies with some rules in place masked it first: they rewrote fields such as name and date of birth with a random function before loading them. More companies skipped that step and loaded the data as it was.

I ran into it while helping clients with architecture and business systems too. Convenience-store members' purchase histories and shopping-mall purchase records were handled carelessly; people loaded production data straight onto development servers.

How it was judged then. The reason I heard most often: access inside the internal network was barely restricted anyway, so it made no difference whether the data sat on a production server or a development server. The industry broadly thought this was fine.

How I see it now (my judgment). It is deeply unprofessional. A copy in a test environment is a full copy of production data, only with weaker protection. Worse, once production data has been accessed, you often cannot tell what was read, so you have to treat the maximum count as fully leaked.

We are a small company and do not outsource development. Only specific people can reach our production servers, and nearly all access is logged. Once a company passes a certain size and relies heavily on outsourcing, this gets much harder.

Since October I have been upgrading all kinds of software inside the company, including code nobody had maintained for years. I used to think a small tool site could go unmaintained. Recently I found that is not acceptable, because it may connect from the outside to the inside. Our approach now is standardization: AI writes the code and AI reviews it, and we use the same technology, version and framework for the same job wherever possible, so one scan tells us whether known vulnerabilities are present.

Other incidents in the same month: four root causes

The large leaks of this month cannot be blamed on "permission settings" alone. There are four kinds of root cause: data kept where it should not be, outsourcing and SaaS supply chains, known vulnerabilities left unfixed, and credential stuffing.

IncidentAnnounced (2026)Scale (as announced)Root cause
Mercari: Fireworks AI, a processor behind Notion AI, breachedNov 203,796 people, including 1,522 customersSupply chain (AI tool)
MUFG Bank: KARTE breachedNov 18About 6.5 million records (gender, age, region)Supply chain (marketing SaaS)
mineo: KARTE breachedNov 18Up to 1.42 million recordsSupply chain (marketing SaaS)
Rakuten SecuritiesNov 17Up to about 15.72 million accountsData in the wrong place (test environment)
MitsumoreNov 16Up to 366,850 recordsCredential stuffing + feature vulnerability
SMBC: email vendor ARARA breachedNov 5227,216 email addressesSupply chain (email delivery)
GMO: Onamae.com Desktop CloudPress conference Nov 2FX accounts manipulated; compensation and related costs estimated at ¥3.05 billionUnfixed vulnerability (flagged in April 2024)
MUJI: MUJI appOct 30 (second report Nov 4)Up to about 27.26 million records; about 1.09 million including full card numbersVulnerability + data in the wrong place (card data that should have been destroyed)

Each category has a cautionary case or a good example:

  • Data in the wrong place. MUJI's president admitted the company kept card data it should have destroyed. Kaitech, a care-staffing service, ended its KARTE contract at the end of January 2025, yet its data stayed in the vendor's system, and 27,530 email addresses were taken this time (Impress Watch).
  • Supply chain. mineo never gave KARTE its payment data, ID documents or login credentials, so none of those leaked. The Mercari case shows that what employees type into AI tools can also flow to processors you never see.
  • Unfixed vulnerabilities. An expert at a GMO group company flagged the problem in April 2024, and it sat unfixed for more than two years. The attacker got in on August 5; GMO confirmed it on October 20.
  • Credential stuffing. Attackers used IDs and passwords obtained elsewhere to try 812 logins at Mitsumore and got into 201 accounts, then used a feature vulnerability to pull more data. Passwords leaked somewhere else become tools to attack you.

Technical fixes you can apply as written

The seven items below are in priority order. Most of the tools are open source or built into cloud platforms, so a small company can adopt them.

1. Keep production personal data out of test environments

  • Block copies of production data to non-production environments by default. Exceptions go through a ticket naming the approver, the purpose and an expiry date; data is deleted automatically at expiry.
  • Mask at copy time, not after the copy. For PostgreSQL, use PostgreSQL Anonymizer or Greenmask. Where synthetic data is enough, use synthetic data (Faker, SDV).
  • Scan non-production environments for real personal data on a schedule: the open-source Microsoft Presidio, Amazon Macie (S3), Google Cloud Sensitive Data Protection.
  • Plant a few fake "canary" customer records in the production database, each with a dedicated email address. Alert when that address receives mail or the record shows up in a non-production database.
  • Put non-production environments in separate AWS accounts or GCP projects with no internet-facing entry point; operate them through SSM Session Manager or IAP.

My judgment: Rakuten Securities had an approval process for exceptions. What it lacked was a hard rule that data must be masked before it is copied. An approval answers "should we do this?" It does not answer "what happens if it leaks afterwards?"

2. Production data: no standing access, query-level auditing

  • Use just-in-time (JIT) elevation: AWS TEAM, GCP Privileged Access Manager or Teleport. Tie each elevation to a ticket number and revoke it automatically at expiry.
  • Turn on pgaudit, RDS audit logs and CloudTrail data events, and keep them for at least a year (my recommendation). Alert on unusual row counts in a single query, bulk SELECTs outside working hours, and sudden jumps in outbound traffic (VPC Flow Logs, GuardDuty).
  • Encrypt fields such as income, assets and withdrawal accounts at the field level. There is a legal benefit as well: personal data protected by measures such as strong encryption does not need to be reported to the Personal Information Protection Commission or notified to individuals if it leaks (BUSINESS LAWYERS).

Why query-level logs matter: without them you cannot prove what was actually read, so you have to report the maximum count.

3. Vendors and SaaS, AI tools included

  • Keep a list of the fields each vendor receives, and send pseudonymous IDs wherever possible. In this incident, what may have leaked from TOYOTA Wallet through KARTE was only about 120,000 internal customer numbers; the number alone cannot identify a person, log in or make a payment (SmartScope).
  • When a contract ends, require proof of deletion and check it yourself.
  • Give each vendor a different seed email address so that you can tell which one leaked.
  • Ban pasting customer personal data into SaaS AI tools. Filter prompts for sensitive data before they leave (Presidio can do this). Process highly sensitive data with a self-hosted private model.

4. An AI code-security scanning pipeline

First, a clarification. Code scanning works on the repository; it needs no connection to any server and no data. The real risk is that source code and secrets in the repository are sent outside. As of October 4, 2026, OpenAI's documentation did not say how long cloned code and scan artifacts are kept (beri.net). So run gitleaks first, including the git history, then rotate every secret it finds and move it out of the repository.

  1. Every PR (free): Semgrep CE, gitleaks, OSV-Scanner or Trivy; add CodeQL for public repositories.
  2. Every release (paid AI): the Codex Security CLI supports --diff main...HEAD, scanning only the changes between two refs (walkthrough). Run a full-repository deep scan only for major releases or once a quarter, with a token budget cap.
  3. Staging (synthetic data): dynamic scanning with OWASP ZAP.

Export all results as SARIF and collect them in GitHub code scanning. Let AI draft the patch; a person reviews it before merge.

Cost reference: when I scanned one repository with OpenAI, it cost about US$50. OpenAI does not publish a per-scan price; billing is by token, so larger repositories cost more. You do not need to scan every day. One scan per major release already fixes a lot.

5. Attack surface and consistent versions

  • List every internet-facing entry point first, including small tool sites nobody maintains. Shut down what you do not use; put what you keep behind Cloudflare Access or IAP.
  • Generate an SBOM with Syft and match it against vulnerabilities with Grype or Trivy; let Renovate open upgrade PRs automatically.
  • Example: on November 18, F5 and nginx.org disclosed 10 NGINX vulnerabilities, fixed in 1.30.7 and 1.31.9. One of the finders, Xint, used an AI testing harness to find four of the critical ones within 90 minutes (eSecurity Planet). Japan's IPA warns that disabling HTTP/2 alone does not cover all of the critical flaws (IPA). Check the nginx bundled in container images and Kubernetes ingress controllers too.
  • The patch deadline I recommend: 72 hours for critical vulnerabilities on internet-facing systems.

My judgment: AI has cut the time to find a vulnerability to minutes. Defenders need the same tools.

6. The user side: preventing follow-on harm

  • Copy what Rakuten Securities got right this time: no URLs in notification emails, a push for passkeys, and a setting that suspends trading and withdrawals (Business+IT).
  • Build credential-stuffing defenses into your own product: login rate limits, a breached-password check through Have I Been Pwned's k-anonymity API, and step-up verification for unusual logins.

7. A breach-response runbook for Japan's personal information law (APPI)

Many people think "you must go public once more than 1,000 people are affected." In fact the law requires a report plus notice to the individuals (Issoh):

  • Four categories require a report: sensitive ("special care-required") personal data, risk of financial harm, acts with improper purpose, and more than 1,000 individuals. Any one category triggers a report, even for a single record. Unauthorized access counts as an improper purpose, regardless of how many people are affected.
  • The duty arises at the "possible leak" stage. The preliminary report is due in roughly 3–5 days and the final report in 30 days, or 60 days for improper-purpose cases, counted from the day any department learns of it.
  • For individuals you cannot reach, public notice plus an inquiry desk can stand in for direct notice.
  • Prepare in advance: a data map (which system holds which fields for how many people, former customers included), templates for the preliminary report, final report and individual notice, and an inquiry desk.

Self-check list

If you cannot answer any one of these 10 questions, start there.

  1. Do test, development or staging environments hold real customers' names, contact details or account data?
  2. When production data is copied to non-production, is masking mandatory, with automatic deletion at expiry?
  3. Who has standing access to the production database? Can it become just-in-time access?
  4. After an incident, can your logs show which tables and how many rows were actually read?
  5. Has data on former customers, and data held by vendors whose contracts ended, been deleted as required?
  6. Which fields does each vendor and SaaS receive? Could they be pseudonymous IDs instead?
  7. Are employees typing customer personal data into AI tools? Who processes it behind those tools?
  8. Have you listed every internet-facing entry point? Are unmaintained tool sites shut down?
  9. Is NGINX on 1.30.7 / 1.31.9 or later, including inside containers and ingress controllers?
  10. If personal data leaks, can you file a preliminary report within 3–5 days?

How we can help

SolanaLink Co., Ltd. helps Japanese SMEs with legacy PHP websites through "stop the bleeding → upgrade → migrate," and hardens the security of crypto exchanges and wallets. If any item on the checklist above is hard to answer, contact us through the form on solanalink.jp.

Notes. Information in this article is current as of November 21, 2026; Rakuten Securities' investigation into the cause is still under way. The statement that production data was copied without masking comes from Rakuten Securities' answers to reporters, not from its formal notice. The legal section is a summary and is not legal advice.

Sources

Comments (0)