AI & Technical question

A frontier lab says existing security benchmarks are too shallow and too easy to game. Design an evaluation product where a task is marked solved only when the exploit reliably reproduces or the patch fixes the issue without breaking intended behavior. What would the task format, execution environment, grader design, and reward/verification logic look like?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Tests whether you can design a rigorous, gameable-resistant evaluation product for security tasks.

Define the task unit: a real or seeded vulnerability in a runnable repo, with the exploit and patch withheld from the model.
Run the exploit in an isolated, network-segmented environment against both the vulnerable build and the candidate patch.
Grade behaviorally: solved only if the exploit reproduces pre-patch, fails post-patch, and existing tests still pass.
Add negative controls, plausible patches that do not truly fix the bug, to catch reward hacking.
Separate partial-credit reward shaping from the final binary verification signal.
Pilot across several CWE classes with human security reviewers before scaling.

Grounding 'solved' in functional reproduction, not diff similarity to a reference patch, is what makes it ungameable.
Denying the model the exploit or test harness, and randomizing environment fingerprints, closes common cheating paths.
Mixing real disclosed CVEs with synthetic seeded bugs avoids contamination from memorized writeups.

Chapters of the AI PM course that teach what this question tests.