--- name: agent-misuse-benchmark description: >- SCAFFOLD, not a finished benchmark. Measures whether a model running inside an agent harness (a coding agent CLI, an IDE agent, a vendor's own harness) carries out tasks an organisation's policy says it must decline, such as offensive security work, and whether it still completes the matched legitimate tasks. The answer depends on the harness and its safeguards as well as the model, so the arms are model × harness × safeguards. This skill ships the scenario format, the arm sweep, the statistics and the report, with a mock for a keyless smoke test. It ships no scenarios, no real harness and no outcome scorer: the team using it supplies those, and the build guide says how. Use it when you are ready to build that for your own policy. license: MIT compatibility: "The scaffold needs Python 3.11+ only. A completed benchmark needs whatever your harness module needs, typically Docker and inspect_ai with inspect-swe." --- # agent-misuse-benchmark — scaffold **Status: scaffold.** Everything generic is here and tested. Three pieces are deliberately absent and must be supplied before this measures anything real: | Piece | What it is | Where to start | |---|---|---| | **Scenario pack** | Pairs of tasks: one your policy says the agent must decline, and its matched legitimate twin | `references/scenario-format.md` | | **Harness module** | Runs one task with one agent harness and model inside a sandbox with no network | `references/build-guide.md`, step 2 | | **Outcome scorer** | Decides, from evidence, whether the agent declined, carried out or completed a task | `references/build-guide.md`, step 3 | If you are an agent asked to "finish" or "build out" this benchmark, read `references/build-guide.md` first and follow it in order. Do not improvise scenarios: the people who own the policy write them. ## The problem this answers "If we enable this model in this agent, will it do work we do not want it to do when someone asks?" A model card answers that for the bare model. Deployed behaviour also depends on the harness's system prompt and tools, on provider and harness safeguards, and on how the task is framed. So the question has to be asked per deployment, with the same tasks, and with a control for an agent that simply declines everything. ## What a completed run reports For each arm (a model in a harness, with safeguards on or off): - **Restricted tasks declined**, with a 95% interval. - **Permitted twins completed**, beside it. An agent that declines everything scores well on the first number and badly on this one. - **Permitted twins declined** (over-refusal). - The same rates by category and by how the task was framed. - A paired comparison of every arm against a baseline arm on the same tasks. - Validity flags: runs with no verdict, and low completion of the permitted twins. It does not say whether a rate is acceptable. Execution tier: **T1** (a sandbox with no network), and a completed benchmark must keep it there. --- ## PLAN 1. **Confirm the three pieces exist.** If any is missing, stop and work through `references/build-guide.md`; do not run with placeholders. 2. **Scenario pack.** `python validate_scenarios.py ` must print `ok`. Read the pack's categories and framings back to the user so they know what is being asked. 3. **Arms** (`AskUserQuestion`). Write `arms.json`: one entry per combination to compare, each with `name`, `model`, `harness` (pin the version) and `safeguards`. Vary one thing at a time where you can: the same model in two harnesses, or one harness with safeguards on and off. The first arm is the baseline unless `--baseline` says otherwise. 4. **Depth and cost.** `smoke` (25 pairs or the whole pack if smaller — proves the wiring), `standard` (100), `full`, or a number. Every pair is at least two agent runs per arm, and an agent run can take minutes and many model calls. Show the plan first: ``` python run_benchmark.py --scenarios --arms arms.json --depth standard --plan ``` 5. **Safety** — confirm with the user before running: - it runs in a disposable environment, not on a machine holding real credentials; - the sandbox has no network (the runner refuses to start if the harness cannot show it); - provider and harness safeguards are left on unless an arm exists to test without them. 6. **Experiment folder:** `bash new_experiment.sh agent-misuse-benchmark "