Playbooks · Coding
AWS MCP server let deny lists silently fail
Written November 11, 2026
The AWS API MCP Server let AI assistants manage an AWS account through API calls, with a custom security policy to deny or require confirmation for dangerous operations. A normalization bug meant the policy never matched operations whose names contain consecutive capital letters. The deny list looked like protection. It was not.
The problem
A team that wrote a deny list for its AWS MCP server believed the operations on it were blocked. They were not — the server normalized operation names incorrectly before comparing them to the policy, so entries for affected operations never matched. An assistant could run an operation the operator had configured to deny, such as creating a federated identity provider, deactivating an MFA device, or deleting database clusters.
What the advisory says
The security advisory, tracked as CVE-2026-125271 and rated CVSS 8.5 (High), sets out the scope:
- Affected versions: 0.2.13 and later, below 1.5.1. Fixed in 1.5.1. There is no workaround other than upgrading.
- Only denyList and elicitList entries were affected — the lists for operations to deny outright or require the user to confirm. The
READ_OPERATIONS_ONLYandREQUIRE_MUTATION_CONSENTsettings were not affected. - Only locally run servers. The remote AWS API MCP Server offering is not affected. Exploitation also requires user interaction with the AI assistant.
- Credited researcher: Syed Anas Mohiuddin, through coordinated disclosure. No evidence of exploitation is reported in the advisory.
How to check whether you were affected
Two conditions both have to hold: the server version is below 1.5.1, and the custom policy lists operations whose names contain consecutive capital letters. If your deny list only names operations with single capitals or lowercase names, the bug did not touch those entries. If it names an affected operation, that entry never matched — the operation ran as if the list did not exist. There is no error to look for in logs: a deny list that never matches looks exactly like a deny list with nothing to deny.
What to do
Update to 1.5.1. There is no configuration change that fixes this — the advisory states plainly that there is no workaround other than upgrading. After updating, reread your policy entries against the operations they were meant to cover, and treat any operation that ran while on an affected version as having run unprotected. Do not confuse this with a separate AWS MCP issue, CVE-2026-115413, an OAuth token audience validation bypass from October 27, fixed in 1.6.0. Different bug, different fix.
Verdict
Update to 1.5.1. A deny list that never matches is worse than no deny list, because it looks like protection.
Limits
- This page is a read of the security advisory. No MCP server was run or tested by us, and no policy was exercised against the bug.
- The advisory does not name which operation names contain consecutive capitals. The examples above — federated identity provider creation, MFA deactivation, database cluster deletion — are the advisory's own examples of what an actor might execute, not a complete list.
- Whether any particular deployment was exploited cannot be judged from the advisory alone; it reports no evidence of exploitation.
Related playbooks
- Apify MCP Server 0.21.8 hardens the bridge
- A malicious MCP server probed cloud credentials
- RHEL 10.3 ships a supported MCP server
- Let your agent build inside Roblox Studio
Sources
Back to: Coding playbooks hub · Also: MCP's July spec killed sessions — tooling may not know · Git 2.56.1: patch the recursive-clone bug