← All updates

Security / News explained

An AWS cost-report tool could write beyond its report folder

AWS’s Pricing MCP Server advisory combines a local file-write flaw with an incorrect read-only annotation. The fix confines report paths, and the deprecated cost-analysis package needs migration.

By Zyel · Published · 4 min read

AWS disclosed CVE-2026-126816 on November 16, 2026: the AWS Pricing MCP Server’s cost-report tool could write attacker-controlled content outside its intended output directory. AWS lists awslabs.aws-pricing-mcp-server versions before 1.0.31 and awslabs.cost-analysis-mcp-server through 1.0.5 as affected. [1]

The project advisory places the issue on the machine running the locally installed MCP server. An actor who influences content read by the agent could cause a report to create or overwrite files that the server process can write. The maintainers explicitly state that no managed AWS service is affected. [2]

Our interpretation is that the report’s everyday business purpose made its side effect easy to underestimate. Looking up a price can be a read operation. Saving the resulting report changes the local filesystem. Both can appear within one conversational task, even though the permissions and review needed for each are different.

The read-only label described the wrong behaviour

The affected tool was annotated as read-only. According to the maintainers, MCP clients configured to auto-approve read-only tools could therefore run it without user confirmation. Exploitation still requires the local server and an agent processing actor-influenced content; this is not a claim that every pricing query is an attack. [2]

The MCP maintainers’ March 2026 explanation describes tool annotations as hints, not guarantees. An untrusted server can describe its behaviour inaccurately, and annotations themselves do not enforce filesystem or network restrictions. A client’s confirmation policy and the implementation’s actual limits therefore remain separate controls. [4]

In this case, the mismatch is especially instructive: a trusted-looking description can be wrong without the client being malicious. Our operational takeaway is to connect approval policy to a clear inventory of actual side effects. A reassuring label is useful documentation; it should not be carrying the whole security boundary on its own.

The fix changes both path handling and the annotation

The project advisory says report paths are now resolved inside a base directory, configured through AWS_PRICING_MCP_OUTPUT_DIR and otherwise defaulting to the process working directory. Paths outside that boundary, traversal segments and escaping symlinks are rejected. The report tool’s readOnlyHint has also been changed to False. [2]

Those are two distinct improvements. Path confinement limits where a report can land. Correcting the annotation gives the client more accurate information about the write. Changing a label alone would leave a vulnerable path handler; confining a path while retaining misleading metadata would still leave the approval experience describing the wrong operation.

PyPI’s release records date version 1.0.31 to June 9, 2026. The November announcement documents the security significance of an already available release. It should not be read as a claim that a new fixed binary first appeared on the disclosure date. [3]

For an application owner, the configured report directory is also an operational dependency. A workflow may expect another process to collect those files. Our proposed acceptance outcome is that intended reports still reach their approved location while destinations outside it are refused. This is a review criterion, not a report of a live exploit test.

The older cost-analysis package needs a migration decision

AWS identifies 1.0.31 as the fixed AWS Pricing MCP Server release. Users of the deprecated cost-analysis package should move to the pricing server at 1.0.31 or later; AWS says that older package will not receive a patched release. [1]

For users unable to upgrade immediately, AWS recommends removing generate_cost_report from auto-approval configuration, requiring confirmation and checking the requested output path. It also recommends avoiding untrusted content in the agent’s context. [1]

That temporary review needs a useful question attached to it. A person approving “make a report” may be considering the report’s subject rather than where it will be written. In our view, the destination and the server process’s writable area are central to understanding what is being authorised.

A hypothetical shared workstation illustrates the scope. If a report process can write another application’s configuration, an escaped report path may affect more than a document. A narrowly writable report account has a different reach. These are permission-dependent possibilities, not evidence of arbitrary administrator access or a reported customer incident.

The practical response combines the software change, an explicit migration for deprecated installations and a review of report-writing authority. Cost analysis may begin with numbers from a remote API, but its output can become a lasting local change. The tool’s path rules and its approval description should both acknowledge that.

Sources + further reading

Source dates describe the original material. Publication above is the date this Zyel article went live.

  1. CVE-2026-126816: AWS Pricing MCP Server report-path traversalAWS Security · Checked Nov 17, 2026
  2. GHSA-mqm3-w68j-75mw: Cost-report file writes and tool annotationAWS Labs MCP maintainers · Checked Nov 17, 2026
  3. AWS Pricing MCP Server 1.0.31 package and release historyAWS Labs on PyPI · Checked Nov 17, 2026
  4. Tool Annotations as Risk Vocabulary: What Hints Can and Can’t DoModel Context Protocol maintainers · 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.

10–2,000 characters. Plain text. Leave out passwords, private project data and personal contact details. A blank display name appears as Anonymous.

Complete the CAPTCHA before sending. Verification runs on your device.