Compare

Scripted vs adaptive attackers.

A scripted attacker follows a set plan: the same tactic, pushed the same way, so results are comparable between versions. An adaptive attacker reads the agent’s replies and changes tactic when one fails, as a real customer or fraudster would, so it finds weaknesses a script misses but varies more from run to run. A good test uses scripted attackers for comparison and adaptive ones for realism.

01Side by side

Compared onScripted attackersAdaptive attackers
BehaviourA set plan per characterChanges tactic based on replies
ComparabilityHigh: same pressure each runLower: needs more runs for a range
RealismLike a determined but predictable customerLike a real negotiator or fraudster
FindsKnown weaknesses, reliablyWeaknesses between known tactics
Risk of overfittingHigher if scripts are visibleLower

02When each is the better choice

Scripted attackers

  • Comparing two versions of the agent fairly.
  • Regression checks on every change.
  • Measuring one tactic precisely.

Adaptive attackers

  • Before launch, to see how the agent holds against a thinking opponent.
  • Agents with high authority (refunds or payments without a person).
  • After the agent already passes the scripted set.

03Using both

Use scripted characters as the steady baseline, add adaptive attackers to probe beyond them, and keep the exact scripts private so the agent is improved, not tuned to the test.

04What Duelpoint does

Duelpoint’s library has scripted characters for every area and adaptive attackers (a haggler, a refund seeker, a fraudster and a supplier) that switch tactic when one fails. Adaptive attackers are included in Assessments and available as an add-on to an Agent Check. See the tests.

Questions

Are the attack scripts public?

The kinds of tactic are described publicly; the exact wording is kept private, so an agent cannot be tuned to pass the test without actually becoming safer.

See where your agent leaks money. Tested against your own rules, with the money at stake.

See the testsHow we score