AI security news · 3 min read

A “read-only” AI tool could write files anywhere.

AWS fixed a path traversal in its Pricing MCP Server’s cost report tool. The tool was marked read-only, which can lead AI clients to run it without asking the user.

A “read-only” AI tool could write files anywhere.: HanulSec guide
Key points
  • CVE-2026-126816 is a path traversal and arbitrary file write in the cost report generator.
  • Someone who can influence the AI agent could write files outside the output folder using absolute paths or “..”.
  • The tool was annotated as read-only, which can lead MCP clients to auto-approve it without user confirmation.
  • Fixed in awslabs.aws-pricing-mcp-server 1.0.31. The deprecated cost-analysis-mcp-server has no fix.
A “read-only” AI tool could write files anywhere.: HanulSec AI Security Brief
AWS Pricing MCP Server · CVE-2026-126816 · the one-page brief we shared on LinkedIn.

What happened

AWS Security Bulletin 2026-184, published on 16 November 2026, describes CVE-2026-126816 in the AWS Pricing MCP Server. Its cost report generator did not stop paths from leaving the output directory, so anyone able to influence the AI agent could make it write content elsewhere on disk. AWS also noted that the tool was marked read-only, which can lead MCP clients to approve it automatically. AWS thanked Ryan von Brubeck for the coordinated disclosure.

Who is affected

  • Users of awslabs.aws-pricing-mcp-server before 1.0.31
  • Users of the deprecated awslabs.cost-analysis-mcp-server (1.0.5 and earlier)
  • Teams that let AI agents auto-approve tools marked as read-only or safe

Why it matters

Many teams let agents run “safe” tools without asking. That decision often rests on a label the tool gives itself. If the label is wrong, nobody is asked before something changes. Through prompt injection, an attacker who can put text in front of the agent can then reach that tool.

What to do now

  • Update awslabs.aws-pricing-mcp-server to 1.0.31.
  • Remove the deprecated cost-analysis-mcp-server, or take generate_cost_report off auto-approve.
  • Review which tools your agents run without approval, and why.
  • Run MCP servers with the least file-system and cloud permissions they need.

How HanulSec helps

  • We list every tool your agents can call and which ones run without approval.
  • We test what each tool can really do, beyond its label and description.
  • We help you keep auto-approve only for tools that are truly harmless.
Want us to test yours?

Tell us what you run and we will send a fixed-price proposal within one business day. Request a quote.

Questions

Quick answers.

What does read-only mean for an MCP tool?
It is a hint the tool gives to AI clients, saying it does not change anything. Clients may use it to skip asking the user. It is not enforced, so it must be accurate.
Do I need to stop using MCP servers?
No. Keep them updated, give them the smallest permissions they need, and only auto-approve tools you have checked.
Request a quote

Scope it in two minutes.

Tell us what you need protected. A security engineer, not a salesperson, replies with a fixed-price proposal.

  • Reply within one business day
  • Free 30-minute scoping call
  • Mutual NDA before any detail is shared
  • Fixed price, free retest of fixed findings