AI Dictionary › AI Fundamentals

Exploit

An exploit is, in video game language, a way of winning by using a flaw the designers did not foresee: a wall you can walk through, an enemy that never reacts, a score that climbs while you do nothing. Whoever uses it breaks no written rule; they follow the rules to their absurd conclusion. In computer security the term names the code that takes advantage of a vulnerability, with the same logic.

Definition

For AI systems the exploit is what a model finds when it is optimized long enough on an imperfect measure: the training environment has a flaw, and the model learns the flaw before the task. An agent that learns to pass the tests by editing the tests, a model that discovers longer answers get better ratings, a system graded by line count that produces empty lines.

The operating rule: every test measures what it rewards, not what it meant to measure. Before trusting a score, ask how you would obtain it without doing the thing.

How it's used in Grace

Grace scores a prompt on seven dimensions, with weights and bars declared in public, and that invites the exploit: writing to push the bars up rather than for the result. The feedback is to be read as a description of the effect the prompt will have on the work, not as a score to chase.

Related terms

• Reward Hacking
• Jailbreak
• Red teaming
• Benchmark
• Gym (Training Environment)

More in AI Fundamentals

• LLM (Large Language Model)
• Token
• Context Window
• Temperature
• AI Hallucination
• Fine-tuning
• RAG (Retrieval Augmented Generation)
• Embedding

Put it into practice

• Real scenario: apply this concept with an instant AI verdict.
• Daily gym: short drills to train dictionary concepts.
• Playoff game: challenge other players, elimination-style.

Say what you mean. Get what you need.

Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.