Key takeaways
- KARTE, a customer experience platform from the Tokyo company PLAID, has been completely stopped since November 13 after attackers exploited a vulnerability in one of its servers. Many Japanese companies use it on their websites and apps to track visitors and send them messages.
- PLAID says the attackers stole credentials for its systems and databases, read data out, and took over PLAID staff accounts to reach the admin screen. Confirmed so far: 79,249 KARTE account email addresses, part of the member data customers had loaded into KARTE, and the secret keys and access tokens customers had registered to connect KARTE to other services.
- The first customer to report, the eyewear chain JINS, says data on up to about 18.29 million member accounts leaked.
- The keys are the bigger problem. Each customer may have registered several, one for each service it connected. With them, attackers could reach those other services directly. In our view, a platform that gathers many companies' keys to many systems is high-risk by design, and that is exactly why it needs protection to match.
What KARTE is#
PLAID calls KARTE a customer experience (CX) platform. A company puts KARTE's tag or software kit in its website and app; KARTE then sees what each visitor does, in real time, and lets the company respond to that person with pop-ups, chat, emails and other messages.
To do that well, companies give KARTE more than browsing data:
- They load their own member data, such as member IDs, names, birth dates and addresses, through a function PLAID calls the linking table, so that KARTE knows who each visitor is.
- They connect KARTE to other services they use, such as messaging channels, advertising platforms, data warehouses and notification tools, so that KARTE can send messages, share audience lists or exchange data on their behalf. To do that, they register those services' secret keys and access tokens in KARTE's admin screen.
Those two things are what make KARTE useful. They are also what leaked.
What happened#
| Date and time (JST) | Event |
|---|---|
| Nov 13, early morning | A third party exploits a vulnerability in a PLAID server, steals credentials for its systems and databases, and reads data out of its databases and other stores. With the stolen credentials it takes over PLAID accounts and reaches the admin screen |
| Nov 13, 08:59 | PLAID detects the access and invalidates the stolen credentials |
| Nov 13, from about 14:27 | Servers are stopped and every KARTE service, for every customer, goes into maintenance |
| Nov 13, about 15:23 | Containment complete; PLAID says there has been no new access since |
| Nov 15 | PLAID says the stop is also meant to stop information that was continuing to flow out; begins contacting customers and asks them to invalidate their integration keys |
| Nov 16, evening | PLAID's filing sets out what leaked |
| Nov 17 | JINS reports up to about 18.29 million member accounts. KARTE is still stopped, with no restart date |
What leaked#
| Data | Count | What it is |
|---|---|---|
| Email addresses of KARTE accounts | 79,249 | The logins of customers' staff who use KARTE |
| Credentials for integrations with external services: secret keys, access tokens | Under investigation | Registered by customers to connect KARTE to other services |
| Part of the data customers imported, such as member data loaded through linking tables | Under investigation | Whatever columns each customer chose to load; may include personal data |
JINS shows what the third line can mean. Its member data in KARTE included member IDs, email addresses, surnames, birth dates, ages, genders, postcodes and prefectures for up to about 18.29 million accounts. More customers' notices are likely to follow as PLAID contacts them. We track the large ones on our list of 2026's largest breaches, and PLAID's updates in our marketing SaaS article.
Why the keys matter#
A key that KARTE uses to call another service works just as well for an attacker. PLAID says plainly that if customers do not invalidate them, attackers could get into the connected services and cause unintended writes or leaks there. In other words, the breach of KARTE can become a breach of each customer's messaging account, advertising account or data warehouse.
Three features make this worse than an ordinary leak:
- Many keys per customer. A customer that connected KARTE to several services registered a key for each. The number of keys exposed may be a multiple of the number of customers.
- They cannot be stored like passwords. A password can be kept as a one-way hash, because the system only has to check it. An outbound key has to be used, so the platform must be able to turn it back into its usable form. Encryption helps only as far as the decryption key is kept somewhere the attacker cannot reach.
- The customer has to act. Only the customer can revoke a key in each connected service and check that service's logs for misuse, and only if it knows which keys it gave KARTE.
PLAID has not said where the keys were stored or how they were exposed. In our reading, there are three plausible routes: they were stored encrypted in a database, and the means to decrypt them sat on the same compromised servers, perhaps in the same configuration files as the database credentials; or the taken-over staff accounts could display them in the admin screen; or a separate store of secrets was reached with credentials kept on those servers. The final report should say which.
What we don't know yet#
- The vulnerability, and how the database credentials were obtained
- Where the integration keys were stored, whether they were encrypted, and by which route they were taken
- How many customers' keys and imported data were exposed, and whether any keys have been misused
- What other data KARTE holds, such as visitor behavior histories, and whether it was affected
- When KARTE will restart, and what will change before it does
Our view#
A hub that holds everyone's keys is high-risk by design. KARTE's value comes from connections: to each customer's own member data and to the other services that customer uses. Each connection means one more key held on the customer's behalf, and many customers have several. Put together, the platform holds keys to a large part of its customers' marketing systems in one place. That concentration is not a flaw peculiar to PLAID; it is the design of every platform of this kind. But it means one server vulnerability and one set of stolen credentials can reach far beyond the platform itself.
Because the risk is high, the protection has to match it. In our view, a platform that holds other companies' keys should assume that its application servers and its staff accounts will one day be compromised, because here both were. The keys should therefore not be readable from either:
- Keep integration secrets in a separate vault, with its own access controls, that application servers can ask to use a key but not read the key itself, where the design allows.
- Encrypt each customer's secrets with its own key held in a key management service, not in configuration files on the same servers.
- Make secrets write-only in the admin screen: once saved, they are never shown again, to staff or anyone else.
- Prefer short-lived tokens with narrow permissions over permanent keys wherever the connected service supports it.
- Watch for bulk reads of the secret store and alert on them, as for any crown-jewel data.
Customers carry half of the job. A company that registers its keys in a platform like KARTE should treat that platform as one of its most critical suppliers. Issue each integration its own key, keep a list of which keys went where, and watch the connected services for unusual use. "Only the permissions it needs" becomes concrete once you look at which way each integration runs:
| Integration | Example | Permission it needs | How to limit it |
|---|---|---|---|
| Into the platform (import) | Member data pulled from a data warehouse or storage | Read | Read-only, and only for the dataset or folder the platform actually uses. A leaked key then exposes that data and nothing else |
| Out of the platform (export) | Results or audience lists written to a warehouse or an ad platform | Write | Write-only where the service allows it, to a dedicated dataset or folder. A leaked key can then add data but not read what is already there |
| To your customers (delivery) | Messages sent through your messaging or email accounts | Send | Cannot be read-only, and a leaked key lets an attacker send phishing from your official account. Restrict it to the platform's source IP addresses where the service supports that |
PLAID's warning covers both "unintended writes" and "leaks", which suggests both kinds of keys were among those exposed. Above all, never register an administrator key because it is quicker to set up: one leaked admin key can read, write and send. Load only the member columns the platform actually needs: JINS's case shows how much a "linking table" can hold. And when the platform says "invalidate your keys", the list is what lets you do it in an hour instead of a week.
What this means for practitioners#
- If you use KARTE: invalidate every key and token you registered in it, in each connected service, and check those services' logs since early November 13 for use you did not expect. List the member columns you loaded; that list decides what you must report to the Personal Information Protection Commission.
- If you use any integration hub (marketing platforms, customer data platforms, workflow automation): list the keys you have given it, match each one to the direction of its integration (read-only for imports, write-only for exports, source IP limits for delivery), replace any administrator keys, and make sure you could revoke them all quickly.
- If you run such a platform: assume your application servers and staff accounts can be compromised, and design the key store so that neither is enough to read customers' secrets.
Japanese terms at a glance#
| Japanese | Reading | Meaning |
|---|---|---|
| 外部サービスとの連携 | gaibu sābisu to no renkei | Integration with external services |
| シークレットキー | shīkuretto kī | Secret key |
| アクセストークン | akusesu tōkun | Access token |
| 紐付けテーブル | himozuke tēburu | Linking table (customers' imported data) |
| 管理画面 | kanri gamen | Admin screen |
| 無効化 | mukōka | Invalidation, revocation |
| ご利用企業 | go-riyō kigyō | Customer companies |