Quick Take
- What: CVE-2026-126816, a path traversal vulnerability in the cost report generation tool of AWS Pricing MCP Server before version 1.0.31
- Who is affected: Users running awslabs.aws-pricing-mcp-server versions before 1.0.31 or any version of the deprecated awslabs.cost-analysis-mcp-server package (all versions through 1.0.5)
- Severity: Medium (CVSS 5.5), requires local access and user interaction
- What to do: Upgrade to awslabs.aws-pricing-mcp-server version 1.0.31 or later
- Workaround: Remove generate_cost_report from auto-approve configurations and avoid passing untrusted content to the agent
In plain English
AWS Pricing MCP Server is a tool that helps AI assistants get real-time pricing information for Amazon Web Services. Researchers found a security flaw in the part of the tool that creates cost reports. If an attacker can trick the AI agent into using certain file paths (for example “../” sequences that climb out of a folder, or absolute paths), they could make the tool write files anywhere on your computer instead of just in the intended folder. The tool was also incorrectly marked as read-only, which meant some AI assistants might run it automatically without asking permission first.
What you should do: If you use this tool, you should upgrade to version 1.0.31. If you don’t use AWS Pricing MCP Server or AI coding assistants, you can ignore this issue.
What happened
On November 16, 2026, AWS published a security bulletin disclosing CVE-2026-126816, a path traversal vulnerability in the AWS Pricing MCP Server. The flaw exists in the cost report generation tool (generate_cost_report) in versions before 1.0.31.
The vulnerability allows an actor who can influence the MCP agent to write content to locations outside the intended output directory. This is possible through the use of absolute paths or directory traversal sequences (such as ../) in the report output path parameter.
According to the GitHub security advisory, the tool was incorrectly annotated with a readOnlyHint set to true, which could lead MCP clients configured to auto-approve read-only tools to execute it without user confirmation. The CVSS score is 5.5, with an attack vector classified as local, requiring user interaction but no privileges.
The impact is limited to files that the server process can write on the machine running the MCP server. No managed AWS service is affected. Exploitation requires the server to be installed locally and the agent to process attacker-influenced content.
AWS credited Ryan von Brubeck for collaborating on the issue through the coordinated vulnerability disclosure process.
What you should do
AWS recommends upgrading to awslabs.aws-pricing-mcp-server version 1.0.31 or later. Users of the deprecated awslabs.cost-analysis-mcp-server package (all versions up to and including 1.0.5) should migrate to awslabs.aws-pricing-mcp-server version 1.0.31 or later, as no patched release of the deprecated package will be published.
In version 1.0.31, the output_file parameter is now resolved and confined to a base directory, configurable via the AWS_PRICING_MCP_OUTPUT_DIR environment variable and defaulting to the server process working directory. Absolute paths outside the base directory, “../” traversal segments, and symlinks that escape the base directory are now rejected. The readOnlyHint annotation has been corrected to False.
Users unable to upgrade immediately should remove generate_cost_report from any auto-approve configuration in their MCP client, so that each invocation requires explicit confirmation and the output_file path can be checked before approving. Users should also avoid passing untrusted content to the agent.
What it means
This vulnerability highlights security risks in AI agent tooling, particularly around file system operations and tool permission annotations. The combination of path traversal and incorrect read-only labeling created a scenario where automated systems might execute dangerous operations without user oversight.
The issue affects only local installations where an attacker can influence the content processed by the MCP agent. It does not affect AWS cloud services or managed offerings. The medium severity rating reflects the requirement for local access and user interaction, limiting the attack surface compared to remotely exploitable vulnerabilities.
This is not the first flaw in an AWS Labs MCP server this month: see the AWS API MCP Server policy bypass via consecutive capitals. File writes outside the intended folder also hit the SiYuan MCP tools on Windows. New cases are tracked in our AI Security hub.
Sources
AWS Security Bulletin 2026-184-AWS; GitHub Security Advisory GHSA-mqm3-w68j-75mw (awslabs/mcp); NVD entry CVE-2026-126816; PyPI release awslabs.aws-pricing-mcp-server 1.0.31.



