Security / News explained
A Timestream MCP connection could send its token to the wrong endpoint
AWS’s November 16 disclosure concerns an MCP connector sending its configured InfluxDB token to an attacker-selected destination. The response includes upgrading and rotating configured tokens.
AWS disclosed CVE-2026-126681 on November 16, 2026: its open-source Timestream for InfluxDB MCP Server could send a configured InfluxDB API token to an unintended endpoint. AWS identifies versions at least 0.0.12 and earlier than 0.0.19 as affected, with 0.0.19 containing the fix. [1]
The project advisory describes a prompt-injection route. A remote unauthenticated actor could introduce a crafted InfluxDB URL into the MCP client’s context. The server could connect to that destination and attach its configured token. Exposure could then give the actor access to InfluxDB data within that token’s permissions. [2]
Our interpretation is that the destination and the credential belong to the same security decision. A server may legitimately possess a token for an approved database. That possession should not make the token available for whichever address a conversation happens to supply. The URL field is doing more than giving directions.
The fixed package predates the disclosure
PyPI records show version 0.0.19 was uploaded on July 28, 2026. November 16 is the security-disclosure date, not the date that fixed package first became available. An operator can therefore have an old-looking dependency pin that is either affected or already beyond this particular fix boundary. [3]
The versioned package documentation describes connection parameters as one trust boundary. When a call supplies connection settings, it must supply the required set; server-environment credentials do not fill gaps in a caller-provided configuration. A supplied destination must also match the configured endpoint or an explicitly allowed additional URL. [3]
The documentation calls this a breaking change for clients that previously mixed partial caller settings with environment fallback. Clients need either the server’s complete configured connection or a complete caller-supplied configuration for an approved destination. [3]
For operational review, that means successful access to the intended database is only part of acceptance. Our proposed check also asks whether a different destination is refused and whether incomplete connection settings fail safely. Those are acceptance expectations derived from the documented boundary, not test results claimed for a customer installation.
Updating the connector and replacing the token address different risks
AWS recommends upgrading to the latest version, incorporating the fixes into forks or derivative code, and rotating InfluxDB API tokens configured on affected servers. That last action is part of the vendor’s response, alongside the software update. [1]
The reasoning is straightforward: preventing a future misdirected connection does not take back a token that may already have left the server. A response needs to account for the old credential’s continued usefulness as well as the corrected connection handling. The owner of the connector and the owner of the database credential may be different people.
Consider a hypothetical assistant reading an imported troubleshooting note that mentions another database endpoint. The important question is whether that text can redirect a tool’s existing authority. This example illustrates the advisory’s trust boundary; it does not establish that a particular note, assistant session or organisation was attacked.
The possible consequence also depends on the token’s actual scope. The project advisory ties access to the permissions of the exposed token. A review should preserve that distinction instead of treating every token as an unrestricted database administrator. [2]
Temporary limits leave the update outstanding
AWS says no workaround fully fixes the issue. Until an upgrade is possible, it recommends the least token permissions needed and avoiding use with MCP clients that process untrusted content. [1]
That advice concerns this named connector and its use of InfluxDB credentials. A product inventory that merely says Timestream, InfluxDB or AI assistant is too broad to establish the affected path. The useful record identifies the MCP server package actually running, its connection arrangement and the credentials assigned to it.
In our view, the follow-up should be concrete enough for another operator to understand: which connector was updated, which configured token was replaced, which destinations remain approved, and who owns any remaining exception. A version written into a build file and a restarted fixed process represent different stages of that work.
The disclosure’s practical lesson is the pairing of destination control with credential control. Keeping secrets out of the conversation is valuable, but the tool that holds them must also keep authority tied to an approved destination. For affected installations, the vendor’s update and token-rotation guidance deserve to travel together.
Sources + further reading
Source dates describe the original material. Publication above is the date this Zyel article went live.
- CVE-2026-126681: Timestream for InfluxDB MCP Server endpoint restrictionAWS Security · Checked Nov 17, 2026
- GHSA-2f9r-g4f3-wc2j: Intended endpoints and configured InfluxDB tokensAWS Labs MCP maintainers · Checked Nov 17, 2026
- Timestream for InfluxDB MCP Server 0.0.19 and connection-security documentationAWS Labs on PyPI · Checked Nov 17, 2026
Reader discussion
What would you ask next?
Questions, useful corrections and thoughtful disagreement are welcome. Comments are reviewed before they appear.
No reader comments yet.