CVE-2026-125271: AWS API MCP Server policy bypass via consecutive capitals

A normalization bug in the AWS API MCP Server let AI agents run AWS operations that operators had blocked, when the operation name contained consecutive capital letters.

Christopher Ort
By Christopher Ort · 2 min read
Share
A security operations center with analysts reviewing policy rules on multiple monitors, one screen showing a red alert for a bypassed security rule.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. How it works
  5. What you should do
  6. What it means
  7. Sources
i10X Quick Take

Quick Take

  1. What: CVE-2026-125271, a policy bypass in the AWS API MCP Server that let AI agents run blocked AWS operations when the operation name contained consecutive capital letters.
  2. Who is affected: Operators running awslabs.aws-api-mcp-server versions 0.2.13 up to (but not including) 1.5.1.
  3. What to do: Upgrade to version 1.5.1 or later and re-test any denyList or elicitList entries whose operation names contain consecutive capital letters.
  4. Fixed: Version 1.5.1, released August 10, 2026.
  5. Reported by: Syed Anas Mohiuddin through coordinated disclosure.

In plain English

The AWS API MCP Server is a tool that lets AI assistants run AWS command-line commands. A bug in how it checked operation names meant that if an operation name had two or more capital letters in a row, the server would not match it against the operator’s block list or approval list. So even if an operator wrote a rule to block or require approval for that operation, the AI agent could run it anyway. AWS fixed this in version 1.5.1.

What you should do: If you use this tool, you should upgrade and then check that your rules now work correctly for any operation names with consecutive capitals.

What happened

AWS published bulletin 2026-176 on November 10, 2026 for CVE-2026-125271 in the AWS API MCP Server, the server that lets an AI assistant run AWS CLI commands. Syed Anas Mohiuddin reported the vulnerability through coordinated disclosure.

PyPI dates the fixed release, 1.5.1, to August 10, 2026, three months before the bulletin appeared. PyPI also marks the package as superseded by the official AWS MCP server and points to a migration guide.

How it works

The AWS API MCP Server normalized operation names incorrectly before checking them against the operator’s policy. AWS states that operations with consecutive capital letters never matched a correctly written denyList or elicitList entry.

A name holding consecutive capital letters never matched the operator’s entry for it. The call went through instead of being denied or held for elicitation, and nothing in the exchange marked the rule as skipped.

What you should do

Upgrade awslabs.aws-api-mcp-server to version 1.5.1 or later. AWS says there is no workaround, and that forked or derivative code has to be patched separately.

After upgrading, re-test any denyList or elicitList entry whose operation name contains consecutive capital letters, to confirm the rule now matches.

What it means

This vulnerability allowed AI agents to bypass security policies silently. Operators who had configured deny or elicit rules for AWS operations with consecutive capitals in their names would have believed those rules were active, when in fact the agent could run those operations without restriction.

The affected version range is 0.2.13 up to (but not including) 1.5.1. Some exposed installs may be running past their replacement, as PyPI marks the package as superseded by the official AWS MCP server.

Sources

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. GitGuardian maps 45 agentic pentest frameworks, finds 11 expose consoles by default
  2. CVE-2026-129130: agent-device file:// URL disclosure lets remote callers read host files
  3. Claude Code approval files leak 5,343 live secrets on public GitHub
  4. OpenAI reports reinforcement learning raises chain-of-thought controllability
  5. Agent Incident Ledger: 57 of 66 AI agent incidents had no attacker