AIDEFEND in Action AIA-0123 Validated Research

Framelink Figma MCP: Command Injection in a curl Fallback

Framelink’s third-party Figma MCP server connects AI coding clients to design data. Imperva and the maintainer validated that crafted tool arguments can reach an unsafe curl fallback and execute a command with the server’s privileges. Network calls and local STDIO are separate entries; the direct-call evidence does not establish a complete model-mediated attack.

Remote Code ExecutionRuntime IsolationMCP SecurityAI Infrastructure
4 applicable AIDEFEND defenses
Source: Another Critical RCE Discovered in a Popular MCP Server 
Author: Yohann Sillam
Original article: Oct 7, 2025

Threat Analysis

  • A third-party design bridge is affected: Framelink Figma MCP Server, distributed as figma-developer-mcp, connects coding clients to Figma data through the Model Context Protocol. This is not Figma’s official MCP service. CVE-2025-53967 affects versions 0.6.2 and older; the complete fix is 0.6.3.
  • A failure path crosses into the shell: a failed native fetch enters a curl fallback. URL/header-derived strings reach child_process.exec, where hostile syntax can execute commands with the server’s privileges. Imperva describes network tool calls; the maintainer also demonstrates local get_figma_data calls over STDIO. An mcp-session-id identifies a session but does not authenticate its caller.
  • Preserve the evidence limit: direct tool-call execution is validated. Indirect prompt injection, DNS rebinding and broader post-compromise impacts are discussed as risks; this article does not present them as completed end-to-end attacks or malicious incidents.

Applicable AIDEFEND Defenses (4)

AID-H-003.010
Deployed AI Software Vulnerability Remediation Lifecycle
Very High
The developer-tooling owner must deploy a currently supported figma-developer-mcp build containing the complete fix, historically 0.6.3. Identify the actual package and running process behind each IDE, npx cache, network listener and automation launcher; an edited manifest does not restart an old process. Safely exercise the fetch-failure fallback with an inert command marker and a legitimate Figma request. The August partial fix is not equivalent to the September complete fix.
AID-H-034.001
MCP Server Deployment Profile, Transport & Exposure Baseline
High
Declare and verify the real MCP deployment profile: use local STDIO when a network listener is unnecessary; otherwise keep the actual HTTP/SSE listener private and authenticated through a transport-compatible integration. Read back bound IPv4/IPv6 interfaces, proxy routes and authentication results rather than assuming localhost. A session ID is not authentication. This removes unintended remote entry, while authorized or local malicious tool calls still require the code fix. Place the complete MCP service and child processes in a constrained non-root runtime with read-only code, bounded scratch, dropped capabilities, no host home/SSH/socket mounts and externally enforced egress only to required Figma routes. Verify host-write, escalation, credential, process and denied-network probes on every replica. This limits shell-command consequences; it does not eliminate execution inside the service or protect credentials intentionally granted to it.
AID-H-034.003
Server-Side Tool Invocation Validation & Object-Level Authorization
High
The MCP maintainer or a controlled server fork can validate tool arguments at the server dispatcher with strict schemas, bounded strings and semantic checks for expected Figma file/node identifiers and output formats. Reject command-bearing values before the handler; keep errors free of raw payloads. Derive permitted identifier formats from the actual API contract and test legitimate examples, rather than copying a customer-ID schema. This requires server integration and does not replace removal of unsafe shell construction.
AID-E-002
AI Runtime Process, Lease & Pod Eviction
High
After preserving evidence and isolating the affected runtime, an authorized managed-Linux administrator can terminate the affected Framelink server and spawned shell processes. Bind the approved population to host, PID namespace, start time, executable identity and cgroup membership; use pidfd-based termination, disable relaunch, and obtain separately credentialed complete-process absence checks. A successful kill command or a missing process name is insufficient. A root-compromised host cannot prove its own cleanliness: use a trusted external/recovery vantage or leave the result unresolved. This response still requires separate patching, artifact cleanup and revocation of any exposed credentials.

What Defenders Should Do Now

  • Deploy a currently supported release containing the complete fix. The August 13 partial change was followed by the full 0.6.3 fix on September 29, 2025. Inventory IDE launch settings, npx caches, network services and running processes, then test the failing-fetch path with a harmless marker and a valid Figma request.
  • Identify the actual transport and listener. Prefer local STDIO when remote access is unnecessary; otherwise use a private, authenticated transport-compatible gateway and verify IPv4/IPv6 bindings and bypass paths. A modern Streamable HTTP example is not automatically compatible with the legacy SSE deployment.
  • At the server dispatcher, strictly validate Figma tool arguments using the real API contract, reject unsafe formats before handler execution and check all transports. Client-side descriptions, skill approvals and argument schemas alone are bypassable by direct callers. Remove unsafe shell construction as the primary repair.
  • Run the complete MCP service with minimal host access, read-only code, bounded scratch, constrained resources and externally enforced Figma-only upstream routes as appropriate. Verify rejected host writes and unapproved connections from the actual runtime. If execution is suspected, preserve evidence, isolate the environment and assess accessible credentials and host changes before restoring service.

Conclusion

This case fits precise MCP-server controls: verified patch rollout, a deployment profile proven by actual listener/authentication checks, strict server-side argument validation and a constrained runtime. The direct exploit evidence supports those controls without inventing a prompt-injection chain. If native execution has already occurred, exact-process termination on a supported managed Linux host is a separate response step; patching or revoking a credential does not stop that execution.

Last updated: