AIDEFEND in Action AIA-0143 Validated Research

AWS API MCP: Operation-Name Normalization Bypasses Custom Policy

A normalization error in the locally run AWS API MCP Server could cause custom deny and approval lists to miss AWS operations containing consecutive capital letters. AWS IAM remained a separate authorization boundary. The fix is to upgrade the server and verify the effective local policy with the actual operation names.

Permission BypassTool AuthorizationMCP SecurityCloud IAM
4 applicable AIDEFEND defenses

Threat Analysis

  • A mismatch before authorization: The local server converts operation names before comparing them with custom security policies. Incorrect handling of consecutive capital letters meant the name could fail to match an otherwise correctly written denyList or elicitList entry. A user interacting with an assistant could therefore reach an operation the local policy intended to stop or pause for confirmation.
  • Two independent gates: Bypassing this local rule does not grant AWS permissions. The operation still runs under the configured AWS identity and IAM authorization. The remote server offering, READ_OPERATIONS_ONLY and REQUIRE_MUTATION_CONSENT are outside this specific defect.
  • Evidence-bounded impact: CVE-2026-125271 affects local package versions from 0.2.13 inclusive to 1.5.1 exclusive. The disclosure names potentially sensitive operations but does not establish an actual customer intrusion. AWS says upgrading is the only workaround. Supporting primary records: the detailed maintainer advisory.

Applicable AIDEFEND Defenses (4)

AID-H-003.010
Deployed AI Software Vulnerability Remediation Lifecycle
Very High
Upgrade every affected local server and derivative to a verified release containing the 1.5.1 repair. Read back the executable’s resolved package version rather than trusting an updated configuration file. Run harmless test operations in an isolated AWS account to prove consecutive-capital names receive the intended deny or elicitation decision before any side effect.
AID-H-034.003
Server-Side Tool Invocation Validation & Object-Level Authorization
Very High
The server maintainer must enforce authorization inside the final MCP handler using the same canonical AWS operation that the SDK will execute. Preserve the caller’s approved context and require the configured deny/elicitation decision for that exact operation before dispatch. Test equivalent operation-name forms through the public handler, including direct calls that bypass a cooperative assistant.
AID-H-017.002
Least-Privilege Tool Architecture
High
Give the MCP process only the AWS identity and service/resource permissions its approved task requires. An IAM denial remains effective even if local policy matching fails, so test both the tool path and a direct AWS call using that identity. Omit identity-management and destructive operations unless the task actually needs them; inspect effective policies and other attached grants.
AID-D-003.003
Agentic Tool Use & Action Policy Monitoring
Medium
After fixing the gate, collect authoritative allow, deny and elicitation outcomes with canonical operation, policy version and call identity. Alert on repeated rejected high-risk calls and missing decision records. This requires server-side instrumentation and a protected audit sink; counting only denial events from the vulnerable normalizer would miss operations it incorrectly allowed.

What Defenders Should Do Now

  • Upgrade to a fixed release containing 1.5.1 and check every local installation and fork. Confirm the running artifact, then exercise the custom rules with safe operation fixtures in an authorized test account.
  • Separate the local policy decision from AWS IAM. Restrict the MCP identity’s service/resource authority and verify that direct API calls cannot exceed it; an assistant’s agreement to obey a prompt is not backend authorization.
  • At the server boundary, match policy and dispatch against one canonical operation identity. Instrument complete decision records and test deny/confirmation behavior, while keeping monitoring separate from enforcement.

Conclusion

AIDEFEND  connects this defect to verified patch rollout, least-privilege tool authority, final MCP-handler authorization and trustworthy decision monitoring. Each has a distinct role: repair the name mismatch, bound the AWS authority, enforce the real operation and reveal subsequent policy violations.

Last updated: