AWS MCP server flaw could leak InfluxDB tokens via prompt injection
AWS Labs has published a fix for its Timestream for InfluxDB MCP Server, and told users to rotate any API tokens configured on the server.
By The Agentic Times ·

A flaw in the AWS Labs Timestream for InfluxDB MCP Server could allow a remote, unauthenticated attacker to steal the InfluxDB API token held by the server, according to an entry published in the US National Vulnerability Database on 16 November 2026.
The issue is tracked as CVE-2026-126681 and affects the package awslabs.timestream-for-influxdb-mcp-server before version 0.0.19. MCP, the Model Context Protocol, is the interface that lets AI assistants call external tools and data sources. This particular server is the bridge between an AI client and an InfluxDB time-series database.
According to the NVD record, the root cause is "improper restriction of intended endpoints in the connection configuration handling". In plain terms, the server did not properly limit which addresses it would talk to when configuring a database connection. An attacker could introduce a malicious InfluxDB URL into the MCP client's context, and the server would send the configured InfluxDB API token to that attacker-controlled endpoint.
The delivery route is prompt injection, in which hostile instructions are smuggled into text that an AI model reads and treats as a command. Because the attacker only needs to get content in front of the model, the advisory describes the actor as remote and unauthenticated. No credentials on the server itself are required.
The practical consequence is credential disclosure. An InfluxDB API token typically grants access to the database it is issued for, so an exposed token is a direct route into the data the MCP server was built to query. The NVD entry does not describe any exploitation in the wild, and does not say who discovered the issue.
AWS Labs' remediation advice has two parts. Users should upgrade to version 0.0.19 or later, and they should rotate any InfluxDB API tokens configured for the server. The second step matters: upgrading closes the leak path, but it does not invalidate a token that may already have been sent elsewhere. Version 0.0.19 of the package is published on PyPI, the Python package index.
The vulnerability is covered by an AWS security bulletin and by a GitHub security advisory, GHSA-2f9r-g4f3-wc2j, filed against the awslabs/mcp repository. Both are listed as references on the NVD record.
The case is a clear example of a pattern that keeps recurring in MCP server design. The server holds a long-lived secret on behalf of the user. It also accepts configuration values that originate, directly or indirectly, from model context that an attacker may be able to influence. If those two things meet without a strict allowlist of permitted destinations, text in a document or a web page can become an instruction to exfiltrate the secret.
Operators running this MCP server should check the installed version, move to 0.0.19 or later, and treat any token that was configured on a vulnerable instance as potentially compromised. Rotating a token is cheap compared with assuming it was never sent anywhere.
Sources
- nvd.nist.govhttps://nvd.nist.gov/vuln/detail/CVE-2026-126681
