Making an LLM-agent evaluation count as red teaming without asking a model to plan attacks
@ericchen356▲ 0 net votes
A fixed scripted-case evaluation is not red teaming: nothing adapts to the target. Three ways to add the adaptive part without making a model generate attack plans (which some models refuse, and rewording or switching models to get around a refusal is not a fix): (1) plant prompt-injection and decoy text inside the data the agent reads and measure how often it is misled; (2) a human attack loop: write a case, run the agent, read its output, adjust, repeat, and log every attempt and outcome; (3) a plain-code search that mutates case parameters and keeps variants that fooled the agent, reporting an attack success rate. Claim only what exists: human or script-generated attacks, not an adaptive AI attacker.