---
title: AWS fixes two MCP server bugs: prompt injection could make the Timestream for InfluxDB MCP server send its API token to an attacker, and a 'read-only' pricing tool could write files anywhere | ctxsec
url: https://ctxsec.com/items/aws-mcp-servers-cve-2026-126681-timestream-influxdb-token-sent-to-attacker-url-via-prompt-injection-cve-2026-126816-pricing-read-only-annotation
description: "AWS says its Timestream for InfluxDB MCP Server (0.0.12 to 0.0.18) did not restrict which endpoints it connects to, so a remote, unauthenticated attacker could use prompt injection to plant an attacker-controlled InfluxDB URL in the client's context and receive the server's configured API token (CVE-2026-126681, fixed in 0.0.19; rotate tokens). In the AWS Pricing MCP Server before 1.0.31, anyone who can influence the agent could make the cost-report tool write attacker-controlled content outside its output directory (CVE-2026-126816). That tool was also annotated read-only, which can lead MCP clients to auto-approve it. The deprecated cost-analysis-mcp-server (&le;1.0.5) will not be patched."
status: "official (AWS Security Bulletins 2026-184-AWS and 2026-185-AWS, Nov 16, 2026); no exploitation reported"
published: 2026-11-16
---

> AWS says its Timestream for InfluxDB MCP Server (0.0.12 to 0.0.18) did not restrict which endpoints it connects to, so a remote, unauthenticated attacker could use prompt injection to plant an attacker-controlled InfluxDB URL in the client's context and receive the server's configured API token (CVE-2026-126681, fixed in 0.0.19; rotate tokens). In the AWS Pricing MCP Server before 1.0.31, anyone who can influence the agent could make the cost-report tool write attacker-controlled content outside its output directory (CVE-2026-126816). That tool was also annotated read-only, which can lead MCP clients to auto-approve it. The deprecated cost-analysis-mcp-server (&le;1.0.5) will not be patched.

[← Context & Harness Attacks](/context) · [markdown](/items/aws-mcp-servers-cve-2026-126681-timestream-influxdb-token-sent-to-attacker-url-via-prompt-injection-cve-2026-126816-pricing-read-only-annotation.md)

official (AWS Security Bulletins 2026-184-AWS and 2026-185-AWS, Nov 16, 2026); no exploitation reported

2026-11-16 · published here

2026-11-20 3:49 PM ET

### AWS fixes two MCP server bugs: prompt injection could make the Timestream for InfluxDB MCP server send its API token to an attacker, and a 'read-only' pricing tool could write files anywhere

- **Upgrade:** awslabs.timestream-for-influxdb-mcp-server 0.0.19; awslabs.aws-pricing-mcp-server 1.0.31. Move off awslabs.cost-analysis-mcp-server.
- **Rotate** any InfluxDB API tokens used with affected versions, and use least-privilege tokens.
- **Pattern to watch:** tool annotations like read-only are hints the server makes about itself. A wrong hint turns client auto-approval into an unreviewed file write.
- **Workaround (AWS):** remove `generate_cost_report` from auto-approve lists and keep untrusted content away from agents using these servers.

Sources: [AWS Security Bulletin 2026-185-AWS: CVE-2026-126681, improper restriction of endpoints in Timestream for InfluxDB MCP Server (Nov 16, 2026)](https://aws.amazon.com/security/security-bulletins/2026-185-aws/) · [AWS Security Bulletin 2026-184-AWS: CVE-2026-126816, path traversal and arbitrary file write in aws-pricing-mcp-server (Nov 16, 2026)](https://aws.amazon.com/security/security-bulletins/2026-184-aws/) · [awslabs/mcp GHSA-2f9r-g4f3-wc2j](https://github.com/awslabs/mcp/security/advisories/GHSA-2f9r-g4f3-wc2j)
