inspect-rogue-arena

Inspect sandbox provider for running AI model evals on Rogue Arena cyber ranges. You write an ordinary Inspect task; each sample claims a model-eval copy, runs on its own range, and Rogue Arena scores the eval's hops when the trial closes.

• The model runs on your machine, through your own model-provider key (for example Anthropic, Google, OpenAI or OpenRouter).
• The target range runs in Rogue Arena, reached with a personal token.
• Scoring is done by Rogue Arena (it checks each hop); the package never computes a score or submits a key.

Support for running Rogue Arena evals from Harbor is in development.

Quickstart

1. Get access. Running evals needs a Rogue Arena account on a team with model-eval access. Access is not self-service: to get access to our library of ready-built cyber-range eval scenarios and model-eval runs, email info@roguelabs.io. Once you're added, create your token below. You will also need an API key for a model provider (for example Anthropic, Google, OpenAI or OpenRouter).

2. Install (Python 3.10+), with the extra for your provider:

   pip install "inspect-rogue-arena[<provider>]"
   # <provider> is one of: anthropic, google, openai, openrouter
   # or [all-providers] for all of them
   

   The extra pulls in that provider's client library.

3. Create your token. Sign in at https://arena.roguelabs.io and go to Team Hub → Model Testing → Tokens (/team-hub/model-testing/tokens). Click Create token, give it a Token label (e.g. laptop or ci-runner), choose Expires in (7, 30, 90 or 365 days), and submit. The token (ra_eval_…) is shown once: "Copy it now. You won't see it again." It runs any published eval you can deploy, as you.

4. Set your environment:

   export ROGUE_ARENA_BASE_URL=https://arena.roguelabs.io
   export ROGUE_ARENA_TOKEN=ra_eval_…     # the token you just copied
   # plus your model-provider key, e.g. ANTHROPIC_API_KEY, OPENAI_API_KEY,
   # GOOGLE_API_KEY or OPENROUTER_API_KEY
   

   Keep the token in the environment: it is never a config field, so it never lands in an Inspect log, and the CLI redacts it from errors.

5. Check the token and list evals:

   rogue-arena-inspect whoami   # your user/org, token label and last 4
   rogue-arena-inspect evals    # the evals this token can run
   

6. Run the built-in example eval against a model of your choice (replace the id with one from evals):

   rogue-arena-inspect launch --eval <environmentId> --model <provider>/<model>
   

   Pick a model by its Inspect provider id; see Inspect's model providers for the ids and the keys each needs. For a dry run with no model spend, use Inspect's built-in offline mock, mockllm/model.

   The agent's shell tool runs long commands to completion (a background exec the sandbox polls), and the model can ask for a longer timeout per command, so a slow scan, build or download is not cut off.

7. See results. The run prints a log location; open it with Inspect's viewer:

   inspect view
   

   Your runs also appear in the web app under Team Hub → Model Testing → Runs (/team-hub/model-testing/runs), with the per-hop score. From the CLI, rogue-arena-inspect runs lists your runs and rogue-arena-inspect run <id> shows one run's status, score and per-stage result.

8. Revoke a token any time from the same Tokens page: click Revoke on its row. If you lose a token, revoke it and create a new one.

Troubleshooting

• 401 / "set a current token". The token is wrong, expired or revoked. Create a fresh one on the Tokens page and re-export ROGUE_ARENA_TOKEN.
• evals lists nothing. Your team has no model-eval access or no eval shared with it yet. Access is not self-service: email info@roguelabs.io to get access to our library of ready-built cyber-range eval scenarios and model-eval runs, and to have an eval shared with your team.
• A provider error (401/403, a model/region policy block). The model runs through your provider credentials, so this comes from the provider, not from Rogue Arena. Check the provider API key and that the model is permitted for your account and region. launch prints a clean line pointing here instead of a raw provider stack trace.
• A trial won't start (TrialReadyTimeout, or a claim refused with 409). No warm copy and no free deployment/concurrency slot right now. evals shows each eval's launch readiness (ready now (N warm) vs on demand, ~X min); wait for a build, lower concurrency, or retry when a slot frees. A 402 funding denial means the eval could not be funded for your organization.

What the client sends to Rogue Arena

The model's conversation (prompts, reasoning, replies) stays on your machine; what it does on the range (commands and their output, file transfers, team chat posts) passes through Rogue Arena. In detail:

• Run and task identity, on each claim. The eval you chose (its environment id, and the eval's name), your Inspect task name, a run id the client generates, and Inspect's sample id; plus Inspect's epoch, sample uuid, eval id and eval-set id when Inspect sets them.
• Commands and their output. Every shell command the agent runs on the range goes to Rogue Arena to run on the VM, and its stdout/stderr come back through Rogue Arena (which returns at most the last 10 MiB of a command's output). Files you upload or download pass through Rogue Arena as their raw bytes.
• Team chat and activity (team runs only). Chat messages members post, and the activity feed. The feed stores only a summary of each command (the command, its exit code and about 400 characters of output); the command's full output still passes through Rogue Arena when it runs.
• Not sent. Your model-provider key never goes to Rogue Arena, and neither does the model's conversation. Your Rogue Arena token is sent only as the bearer credential.

The rest of this document is reference material.

Authenticate (detail)

export ROGUE_ARENA_BASE_URL=https://arena.roguelabs.io
export ROGUE_ARENA_TOKEN=ra_eval_…

ROGUE_ARENA_TOKEN is your personal model-eval token (ra_eval_…), sent as the bearer on every call. Keep it in the environment: it is not a sandbox config field, so it is never serialized into an Inspect log, and the CLI never prints it (it is redacted from error messages).

Supported Inspect versions: >=0.3.260,<0.3.274 (pip resolves one automatically). 0.3.260 is the first with the sandbox errors the provider raises.

Check the token and pick an eval

rogue-arena-inspect whoami   # userMainId, organizationId, tokenLabel, tokenLast4
rogue-arena-inspect evals    # the evals this token can run, with launch visibility

evals prints one block per eval with its name, version, stage (hop) count, scoring mode, and launch visibility:

example-ad-chain  |  v3 · 4 stages · sequential · ready now (2 warm)
    id: 11111111-1111-4111-8111-111111111111
example-web-app  |  v1 · 2 stages · independent · on demand, ~7 min
    id: 22222222-2222-4222-8222-222222222222

ready now (N warm) means warm copies are claimable now; on demand, ~X min means a claim builds a fresh copy; unknown means readiness was not reported (the command still succeeds). evals --json prints the raw listing; whoami prints JSON. --base-url and --token override the environment, but prefer the environment variable for the token (a flag is visible in the process list).

Runs

rogue-arena-inspect runs             # this token's recent runs (jobs)
rogue-arena-inspect run <id>         # one run's status, score and per-stage result
rogue-arena-inspect run <id> --watch # live-tail the run's events until it settles
rogue-arena-inspect run <id> --chat  # print the run's team chatroom transcript

run <id> takes a run (job) id (or a trial id as a fallback) and prints the run's status and, per trial, the server score and each stage (#<step> <name>: pass/fail). Each trial's terminal reason reads as the Rogue Arena enum plus a short gloss, e.g. reason=timeout (ran out of wall clock).

• --watch polls the run and its trial events on an interval (--interval, default 3 s), printing new exec/hop/lifecycle events as they arrive until the job settles, then prints the final render. A team run is one trial, so it follows that trial like any other.
• --chat prints the run's team chatroom, one line per message.

runs, run, and run --chat take --json for the raw objects.

Launch

rogue-arena-inspect launch --eval <id> --model <provider>/<model>  # built-in example task
rogue-arena-inspect launch --eval <id> --task my_task.py --model … # your own task
rogue-arena-inspect launch --team team.yaml --eval <id>            # a multi-agent team
rogue-arena-inspect launch --eval <id> --model … --max-runtime 1800 # per-trial wall clock (s)

launch is a thin wrapper around inspect eval. It resolves the built-in example task (shipped inside the package), sets ROGUE_ARENA_EVAL, keeps the token in the environment (never on the child argv), and runs inspect eval. Anything after -- is passed straight through to inspect eval (e.g. -- --limit 1). --max-runtime SECONDS sets the per-trial wall clock (60 to 86400).

Provider errors

The model runs on your machine through your own provider credentials, so a provider refusal (a model/region policy block, a bad or missing API key) comes from the provider, not from Rogue Arena. When launch's inspect eval child exits on one of these, launch prints a clean, actionable line instead of a raw provider stack trace; the child's own output is still shown. Check your provider API key, that the model is allowed for your account and region, and that the matching provider extra is installed. Inspect's model providers page lists each provider's model ids and keys.

Write your own task

The example task files ship inside the package, under inspect_rogue_arena/examples/ (eval_example.py, team_example.py, team.yaml, and the eval-example.eval.yaml run config). rogue-arena-inspect launch runs them without a checkout. To customize one, find the folder, copy the file out, and pass it with --task:

python -c "import importlib.resources as r; print(r.files('inspect_rogue_arena') / 'examples')"

Or write your own task. Use the scorer and give the sandbox a config:

from inspect_ai import Task, task
from inspect_ai.agent import react
from inspect_ai.dataset import Sample
from inspect_ai.util import SandboxEnvironmentSpec

from inspect_rogue_arena import RogueArenaSandboxConfig, rogue_arena_hops, rogue_bash


@task
def my_eval() -> Task:
    return Task(
        dataset=[Sample(input="…your objective…", target="scored by Rogue Arena")],
        solver=react(tools=[rogue_bash()]),
        scorer=rogue_arena_hops(),
        sandbox=SandboxEnvironmentSpec(
            "rogue-arena", RogueArenaSandboxConfig(eval="<environmentId>")
        ),
    )

rogue_bash() is the shell tool the examples use: it states its default timeout to the model and runs long commands as a background exec, so a slow command runs to completion and a transient network failure on a poll is retried rather than ending the run.

Teams (multi-agent)

A team run puts several agents on one range. Each member has its own model, system prompt and shell channel into the one VM; they share a chatroom (post_message / read_messages), a scoreboard (post_finding / read_scoreboard) and, in shared_sight, an activity feed. A team is one Inspect sample, so it claims exactly one trial and drives one VM.

Describe the team in a YAML/JSON file (the packaged team.yaml is a template; copy it out to edit) and launch it:

rogue-arena-inspect launch --team team.yaml --eval <environmentId>

The team file sets turn_taking (coordinator, where a lead agent delegates via Inspect's handoff, or round_robin), visibility, optional tool_timeout_seconds, rounds / turns_per_member, and the members (each with model, name, role, prompt).

Command timeouts. A member's bash runs every command in background mode: it runs to completion and returns its full output, for at most tool_timeout_seconds (a top-level team field, default 600), and the tool description tells the model so. The model may pass its own timeout (seconds) per command. A command that outlives its timeout is stopped, with everything it started, and ends with Inspect's timeout error.

Visibility: shared_sight vs communicate_only

visibility decides whether members can read each other's actions. In both modes the activity feed is recorded and sent to Rogue Arena (the run owner can see it on the run page). The feed stores only a summary of each command (the command, its exit code and about 400 characters of output); the command's full output still passes through Rogue Arena when it runs.

• shared_sight (the default) gives each member a read_activity tool to read the feed, so teammates can avoid colliding.
• communicate_only withholds read_activity, and Rogue Arena also refuses a member's read of the feed during the run, so members coordinate through the chatroom alone. The feed is still recorded for the run owner.

Either way, members share the one VM, chatroom and scoreboard. If Rogue Arena does not serve the activity routes, a shared_sight team prints one line and runs chat-only rather than letting a read_activity call end the run.

Team scoring. A team run is one trial; its score is the team's. Rogue Arena scores it when the trial closes, like any other trial, and a key-file hop counts a key any member surfaced.

Sandbox config

Explicit Inspect config wins over environment variables, which win over the defaults. The config is typed, validated and frozen, and rejects unknown keys.

Lifecycle

There is one job per sandbox config per eval run, and one trial per sample:

1. task_init resolves eval and creates the job with n, concurrency, the wall clock and an idempotency key (a transport failure retries once with the same key, so it never creates a second job). Tasks in the same run that share a sandbox config share that job.
2. sample_init waits for a free concurrency slot, then claims a trial with its own idempotency key: a transport failure or a 502/503/504 resends the claim once and Rogue Arena answers the trial it already opened, never a second one. A claim can return a trial still provisioning; the provider polls it to ready within max_wait_seconds, and cancels it if anything fails.
3. Heartbeat. A background task heartbeats the trial every heartbeat_interval_seconds until the wall clock ends; every exec also heartbeats first. Only a 409 (the trial is no longer live) stops the loop.
4. Tool calls exec, upload and download on the trial, each bounded by the trial's remaining wall clock.
5. Scoring. rogue_arena_hops() closes the trial; Rogue Arena scores every hop during the close and returns the scored trial.
6. sample_cleanup closes a trial that is still open (cancelled when the sample was interrupted). A close is sent at most once; if it times out, the provider reads the trial back until it is closed (Rogue Arena finishes the close server-side).
7. task_cleanup closes any leftover trials as cancelled, then cancels the job.

Claim errors

Each of these fails the sample and releases its concurrency slot:

Concurrency

A semaphore per sandbox config, sized to concurrency, holds each claim until an earlier trial closes, so the provider never sends a claim that would be refused (409). Waiting for a slot is ordinary queueing with no time limit.

The provider also implements Inspect's default_concurrency() hook, which sets Inspect's default --max-sandboxes to ROGUE_ARENA_CONCURRENCY (default 2). When a task config sets concurrency, pass the same value as --max-sandboxes so Inspect's queue matches the job's limit.

Commands

exec is synchronous by default. A command whose timeout is at most sync_exec_ceiling_seconds (default 50, at most 55) runs as one request; a command with no timeout gets the ceiling as its timeout.

A longer timeout runs in background mode: the provider launches the command, then polls it (first after 0.25 s, the wait doubling up to poll_interval_seconds) until it finishes, for up to the command's timeout plus 5 minutes (the server ceiling is 4 hours). Rogue Arena stops a background command and everything it started at its timeout and when its trial ends.

A background poll that hits a transient failure retries with exponential backoff (capped at 30 s) until the deadline. Rogue Arena returns at most the last 10 MiB of a command's output. cwd, env and input (stdin) pass through; user, timeout_retry and concurrency are ignored.

Files

write_file and read_file move raw bytes, base64-encoded in JSON. Text is UTF-8. Absolute agent paths only, so a relative path resolves against /. A read_file larger than Inspect's read limit raises OutputLimitExceededError. An upload is capped at 10 MiB.

Scoring: rogue_arena_hops()

The scorer closes the trial and reads the server score the close returns. It never computes a hop locally, never runs a command, and never submits a key.

Key-file hops are proven server-side: Rogue Arena counts a key-file hop only when the planted key appears in the agent's exec output (including background results) or in a file the agent downloads. The agent proves a hop by reading the key through its tools; there is nothing to submit.

• Value: hops_passed / hops_total, as counted under the eval's scoring mode (sequential or independent). The answer reads "<hops_passed>/<hops_total> hops".
• Metadata: trial_id, hops_passed, hops_total, reached_step, scoring_mode, outcome, terminal_reason, timed_out, and per_hop.
• per_hop.passed is the raw check result, not whether the hop counted. In sequential mode a hop after the first failed step does not count even when its check passed; use hops_passed and reached_step for the count.

With Inspect's default fail_on_error, one such infrastructure outcome aborts the whole run. To tolerate them, pass --no-fail-on-error, or a threshold such as --fail-on-error 0.1. From Python, pass fail_on_error=False to inspect_ai.eval(). To retry, --retry-on-error N claims a fresh trial for each retry, so leave headroom in max_trials.

Python client

RogueArenaClient.from_env() reads ROGUE_ARENA_BASE_URL and ROGUE_ARENA_TOKEN. It covers the whole eval API: whoami, list_evals, list_runs, create_job, get_job, cancel_job, list_job_trials, claim_trial, get_trial, heartbeat_trial, run_command (and the single-request exec_trial), upload_file, download_file, close_trial and list_trial_events.

import asyncio

from inspect_rogue_arena import RogueArenaClient


async def show_evals() -> None:
    client = RogueArenaClient.from_env()
    for entry in await client.list_evals():
        print(
            entry.environment_id, entry.name, entry.hop_count, entry.scoring_mode.value
        )


asyncio.run(show_evals())

Development

pip install -e '.[all-providers]'
python -m unittest discover -s tests

The tests run offline: HTTP goes to a mock transport and an in-memory fake server, and the Inspect runs use mockllm, including a run of the shipped example task.

License

Apache License 2.0. See LICENSE and NOTICE. The license grants no rights to the "Rogue Arena" or "Rogue Labs" names or marks.

Changelog

0.4.0

• First public release.