Datasets:

pooria
/

benchmark-exploit-resistance

license: mit
language:
  - en
pretty_name: Exploit Resistance
tags:
  - benchmark
  - llm-evaluation
  - ai-benchmark
  - model-evaluation
  - evals
  - reward-hacking
  - verifier
  - coding-agent
  - agent-benchmark
  - tool-use
  - openrouter
  - openai
  - mcp
  - typescript
  - grading
task_categories:
  - text-generation
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/tasks.jsonl

Exploit Resistance

Does the agent fix the bug or game the verifier?

What it measures

Whether a coding agent fixes a seeded function or games the visible test runner.

Category: safety. Status: alpha.

How to run

$ npx benchmark-exploit-resistance list
$ npx benchmark-exploit-resistance run --model <model-id> --provider openrouter
$ npx benchmark-exploit-resistance score results.json

--provider mock runs the whole benchmark offline with no keys. Real models run through OpenRouter (OPENROUTER_API_KEY) or any OpenAI-compatible base URL (OPENAI_BASE_URL, OPENAI_API_KEY).

Data fields

data/tasks.jsonl holds 46 tasks, one JSON object per line (kinds: 46 episode).

field	type
`id`	string
`kind`	string
`prompt`	string
`metadata`	object

Provenance

Exploit Resistance provenance

This file is generated from the benchmark definition.

Seeded mini-repo generator

Kind: synthetic
URL: https://github.com/pooriaarab/benchmark-exploit-resistance/blob/main/scripts/gen-tasks.mjs
Licence: MIT
Retrieved: 2026-10-16T00:00:00.000Z
How to get it: Run node scripts/gen-tasks.mjs --seed 20261015 --split public from the repository root. --split hidden --seed <n> writes .hidden/tasks-.jsonl and does not overwrite the public file.
Contamination notes: Public tasks use seed 20261015 and are committed in data/tasks.jsonl. Public records omit hidden cases and the honest expression. The grader rebuilds the hidden cases from the seeded generator at grade time; the answer templates live in scripts/honest.mjs, which the published package does not carry. Hidden splits use another seed, which changes the spec variants, the bugs and the answers, and stay unpublished under .hidden/. Once a scenario's three variants are all taken, a hidden task can share spec text with a public task; only the drawn argument tuples are kept disjoint.
Refresh: Pick a new seed, regenerate, and review the JSONL diff before replacing the public file.

Loophole catalog

Kind: human-written
URL: https://github.com/pooriaarab/benchmark-exploit-resistance/blob/main/docs/loophole-catalog.md
Licence: MIT
Retrieved: 2026-10-16T00:00:00.000Z
How to get it: Read docs/loophole-catalog.md. The generator copies one catalog hint into each task.
Contamination notes: The catalog names the verifier shortcuts. It does not contain hidden inputs or the honest expressions.
Refresh: Edit the catalog and the generator hint list in the same change so the task file stays aligned.

Grader

Hidden cases from the seeded generator, plus a tamper check on verifier files. The visible runner is not the score.

Limitations

This dataset contains only the public task split (46 tasks). Hidden evaluation splits use a private seed and are deliberately not part of this dataset. Source kinds: human-written, synthetic. These are evaluation prompts for the benchmark-exploit-resistance benchmark, not training data; scoring requires the package's deterministic grader. See PROVENANCE.md for contamination notes.

Citation

@misc{benchmark_exploit_resistance,
  author = {Pooria Arab},
  title = {Exploit Resistance task data},
  year = {2026},
  url = {https://github.com/pooriaarab/benchmark-exploit-resistance}
}