The best way to properly verify an application or system during a stress test would be to expose it to "live" data that has been sanitized to avoid exposing any sensitive information or Personally Identifiable Data (PII) while in a testing environment. Fabricated test data may not be as varied, complex or computationally demanding as "live" data. A production environment should never be used to test a product, as a production environment is one where the application or system is being put to commercial or operational use. It is a best practice to perform testing in a non-production environment.
Stress testing is carried out to ensure a system can cope with production workloads, but as it may be tested to destruction, a test environment should always be used to avoid damaging the production environment. Hence, testing should never take place in a production environment. If only test data is used, there is no certainty that the system was adequately stress tested.

This means that many of the tools used for ethical hacking have the potential of exploiting vulnerabilities and causing disruption to IT system. It is up to the individuals performing the tests to be familiar with their use and to make sure that no such disruption can happen or at least shoudl be avoided.
The first step before sending even one single packet to the target would be to have a signed agreement with clear rules of engagement and a signed contract. The signed contract explains to the client the associated risks and the client must agree to them before you even send one packet to the target range. This way the client understand that some of
the test could lead to interruption of service or even crash a server. The client signs that he
is aware of such risks and willing to accept them.
The following are incorrect answers:
An organization should use ethical hackers who do not sell auditing, hardware, software,
firewall, hosting, and/or networking services. An ethical hacking firm's independence can
be questioned if they sell security solutions at the same time as doing testing for the same
client. There has to be independance between the judge (the tester) and the accuse (the
client).
Testing should be done remotely to simulate external threats Testing simulating a cracker
from the Internet is often time one of the first test being done, this is to validate perimeter
security. By performing tests remotely, the ethical hacking firm emulates the hacker's
approach more realistically.
Ethical hacking should not involve writing to or modifying the target systems negatively.
Even though ethical hacking should not involve negligence in writing to or modifying the
target systems or reducing its response time, comprehensive penetration testing has to be
performed using the most complete tools available just like a real cracker would.
Reference(s) used for this question:
KRUTZ, Ronald L. & VINES, Russel D., The CISSP Prep Guide: Mastering the Ten
Domains of Computer Security, John Wiley & Sons, 2001, Appendix F: The Case for
Ethical Hacking (page 520).

Explanation/Reference:
Security must be considered in information system design. Experience has shown it is very difficult to implement security measures properly and successfully after a system has been developed, so it should be integrated fully into the system life-cycle process. This includes establishing security policies, understanding the resulting security requirements, participating in the evaluation of security products, and finally in the engineering, design, implementation, and disposal of the system.
Source: STONEBURNER, Gary & al, National Institute of Standards and Technology (NIST), NIST Special Publication 800-27, Engineering Principles for Information Technology Security (A Baseline for Achieving Security), June 2001 (page 7).

A teardrop attack consists of modifying the length and fragmentation offset fields in sequential IP packets so the target system becomes confused and crashes after it receives contradictory instructions on how the fragments are offset on these packets. A SYN attack is when an attacker floods a system with connection requests but does not respond when the target system replies to those requests. A smurf attack is an attack where the attacker spoofs the source IP address in an ICMP ECHO broadcast packet so it seems to have originated at the victim's system, in order to flood it with REPLY packets. A buffer overflow attack occurs when a process receives much more data than expected. Source: KRUTZ, Ronald L. & VINES, Russel D., The CISSP Prep Guide: Mastering the Ten Domains of Computer Security, John Wiley & Sons, 2001, Chapter 3: Telecommunications and Network Security (page 76).

Explanation/Reference:
Before the advent of classless addressing, one could tell the size of a network by the first few bits of an IP address. If the first bit was set to zero (the first byte being from 0 to 127), the address was a class A network. Values from 128 to 191 were used for class B networks whereas values between 192 and 223 were used for class C networks. Class D, with values from 224 to 239 (the first three bits set to one and the fourth to zero), was reserved for IP multicast.
Source: STREBE, Matthew and PERKINS, Charles, Firewalls 24seven, Sybex 2000, Chapter 3: TCP/IP from a Security Viewpoint.