An unchanged default password on a Grafana test instance gave an attacker a route into Chainstack’s production systems, and customer password hashes were among the data the attacker reached. On November 10, the blockchain infrastructure provider added a technical claim to its incident page: it assesses with high confidence that the attacker used ARTEX, an open-source AI agent framework built for automated penetration testing. It did not name who operated the tool.
What Chainstack Disclosed
Chainstack sells managed node access and RPC endpoints for more than 70 blockchain networks, including Ethereum, Solana, and Base. It published its first incident notice on November 3, 2026, then added technical details and indicators of compromise on November 10. The confirmed exposure covers a customer database holding organization names, account-holder names, member email addresses, and password hashes. Chainstack says it does not store full card numbers or security codes, and it has not said how many customers or accounts the breach touched.
Node credentials remain the open question. The attacker accessed the systems that store them, and although Chainstack found no evidence that the credentials were read, it treats them as exposed. It tells customers to change passwords, especially any reused elsewhere, rotate node credentials through the console, regenerate platform API keys if they use the platform API, and watch for phishing. Chainstack says its checks found no modifications to customer nodes, the console, or the API, and that customer-facing services kept running normally.
How the Attacker Reached Production
A Grafana Instance With Default Credentials
The entry point was an internet-exposed Grafana dashboard instance in a development Kubernetes cluster. Chainstack deployed that test instance in July with its default administrator credentials unchanged and no single sign-on or basic authentication in front of it. On October 26, the attacker logged in with those credentials and started using Grafana’s data-source proxy, which had no destination allowlist, to reach internal services.
An MCP Server That Required No Authentication
At 21:01 UTC, about an hour and a half after its first proxy request, the attacker connected to an internal tool server built on the Model Context Protocol (MCP), an open standard for connecting AI agents to tools and data. The server exposed Kubernetes operations to AI agents, required no authentication, and ran with a cluster-admin service account. Chainstack believes the attacker likely used it as a command executor. The company found no evidence that anyone exploited a software vulnerability at any step.
Production Access From October 29
The attacker reached parts of production on October 29 using compromised machine credentials, and read production Kubernetes secrets at 17:49 UTC. On October 31, it bulk-read the production cloud secrets manager, used the private key of a VPN peer to connect to the database network, and gained administrative access to a production PostgreSQL database. It exfiltrated query results and queried the pg_shadow table, which holds database role password hashes, on one instance. Chainstack detected the intrusion that day and cut the attacker’s access paths. It closed the investigation on November 2, after the last observed use of a copied credential at 06:28 UTC.
Attributing the Tool, Not the Actor
Chainstack ties the intrusion to a broader AI-powered attack campaign reported in late October, citing shared tooling, naming markers, and some techniques. It then compared its indicators with the actor-specific indicators CrowdStrike published for Vapor Panda, the China-linked group that CrowdStrike says ran ARTEX mainly on DeepSeek against more than 200 organizations starting October 17. None of Chainstack’s indicators appeared in that set. The company ran the same check against indicators Microsoft published for Storm-0940 and Storm-3358, two other groups Microsoft tracks as ARTEX users, and again found no match. Its conclusion reads, “We therefore attribute the tooling, not the actor.”
The data suggests Chainstack drew the line where its evidence allows. ARTEX is open source, and Microsoft’s reporting describes Storm-0940 as one of many actors across several regions that use it, so finding the framework in an intrusion does not point to any single group. Naming the tool still gives other defenders something concrete to search for: according to Chainstack’s report, CrowdStrike lists artx and artex strings among ARTEX markers.
How Chainalysis and WorkOS Handled the Question
Chainalysis said on October 30 that an external actor used “offensive agentic techniques” to gain access to its systems and internal tools. The notice names no tool, vulnerability, or attacker, and it gives no count of affected customers. WorkOS took a different route. Its November 2 bulletin described an AI-assisted offensive security framework without naming it, though later statements named ARTEX.
On November 6, WorkOS said an unnamed security researcher had claimed responsibility, telling the company that the agent got out of its isolated test environment and reached WorkOS by accident. WorkOS and an independent forensic investigator compared the researcher’s request details, timestamps, and other artifacts with WorkOS’s logs and concluded with high confidence that this person was responsible. Neither party has publicly explained how the agent escaped, and BleepingComputer’s report does not say that anyone verified the escape independently.
Read together, the three notices sit at different levels of specificity. Chainalysis stopped at the technique. WorkOS went further, though its account of how the agent got loose rests on the researcher’s word, and Chainstack stated the one claim its evidence supports and said outright that it was not naming an actor.
Why the MCP Server Matters More Than the AI Label
My take is that the attribution detail, careful as it is, is not the most important finding in Chainstack’s report. The unauthenticated MCP server running with cluster-admin rights is. The Cloud Security Alliance’s AI Safety Initiative flagged the same weakness in a September research note, describing authentication-optional MCP servers as a pattern documented across 2026. In my view, companies wiring AI agents into production systems risk repeating a familiar mistake from earlier waves of internal tooling: they build the powerful tool first and gate access to it later, if at all.
Chainstack found no evidence of an exploited vulnerability, which points to a chain built from configuration choices: default administrator credentials with no single sign-on, a proxy with no destination allowlist, an MCP server with no authentication, a service account with cluster-admin rights, and machine credentials that later opened parts of production. Whether a person, an AI agent, or some combination operated the tool Chainstack identified, the first link was open to anyone who found the Grafana instance. My recommendation for teams that run internal MCP servers is to require authentication on every one, scope their service accounts well below cluster-admin, and keep development environments from holding credentials that open production.
What Comes Next for Breach Disclosures
The signals suggest more companies will face the same choice in the coming weeks: name the tool, name the actor, or say nothing technical. In my view, precision, even partial precision, will earn more trust than a blanket attribution to AI, because a named tool and a set of published indicators give other defenders something to check against their own logs. Security teams can run the practical test today: find the internal tool servers in their own environment and check which ones would answer an attacker’s first request.
The post Chainstack’s Breach Disclosure Names the AI Tool but Not the Attacker appeared first on DataFLOQ.
