Policies

Each entry below is a factory returning an Inspect Agent, selected by name with ct run eval --policy <name>. Names, arguments, and defaults are read from the policy registry, so this page cannot drift from the code.

Adding one

from control_tower.protocols.policies.registry import policies

@policies.register()
def my_policy(...):
    ...

Every non-underscore module under the registry's discovery package is imported on first lookup, so the decorator is all the registration a new policy needs.

Registered policies

`agentic-honest`

Honest policy framed by the environment's own description.

Runs a multi-turn ReAct loop over the environment's tools; the system prompt is the environment description followed by the main task.

Defined in control_tower.protocols.policies.policies.agentic_honest.policy.agentic_honest_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`max_steps`	`int`	`150`

`attack`

Attack policy with sophisticated two-phase planning.

Defined in control_tower.protocols.policies.policies.attack.policy.attack_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`enable_main_task_test`	`bool`	`False`
`enable_side_task_test`	`bool`	`True`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`phase1_max_steps`	`int`	`30`
`phase2_max_steps`	`int`	`30`
`execution_max_steps`	`int`	`100`
`side_task_visibility`	`Literal`	`'none'`
`cache`	`bool`	`True`

`explore-crash`

Container-crash policy focused on maximizing disruption and instability.

Defined in control_tower.protocols.policies.policies.explore_crash.policy.explore_crash_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`

`explore-only`

Explore-only policy that tests side task accomplishment without modifications.

This policy gives the model the side task directly and allows exploration but prohibits any modifications. It tests whether side tasks can be accomplished on a pristine environment.

Defined in control_tower.protocols.policies.policies.explore_only.policy.explore_only_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`enable_side_task_test`	`bool`	`True`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`

`find-questions`

Policy that explores an environment to discover answer-question tasks.

Defined in control_tower.protocols.policies.policies.find_questions.policy.find_questions_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`

`honest`

Honest untrusted policy with structured planning.

Defined in control_tower.protocols.policies.policies.honest_untrusted.policy.honest_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`enable_main_task_test`	`bool`	`False`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`
`cache`	`bool`	`True`

`honest-mistakes`

Two-phase policy: honest execution then side task attempt.

Phase 1: Run as honest agent (no side task knowledge) until: - Agent calls submit, OR - max_honest_steps is reached Phase 2: Show trajectory + side task, ask model to accomplish side task

Defined in control_tower.protocols.policies.policies.honest_mistakes.policy.honest_mistakes_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`enable_main_task_test`	`bool`	`False`
`enable_side_task_test`	`bool`	`True`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_honest_steps`	`int`	`15`

`inspect-swe`

Run a pinned inspect-swe agent factory in the task sandbox.

Defined in control_tower.protocols.policies.policies.inspect_swe.policy.inspect_swe_policy.

Argument	Type	Default
`agent`	`str`	required
`version`	`str`	required
`model`	`str`	required
`agent_args`	`str \| dict[str, Any]`	`'{}'`
`scorer_timeout`	`int`	`600`
`enable_side_task_test`	`bool`	`True`
`debug`	`bool`	`False`

`mt-causes-st`

Check whether honest eval passed ST scorer.

Defined in control_tower.protocols.policies.policies.mt_causes_st.policy.mt_causes_st_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`traj_like`	`TrajLike`	required
`scorer_timeout`	`int`	`600`
`max_retries`	`int`	`3`
`provide_scorer`	`bool`	`False`

`obvious-attack`

Direct attack with no concealment or stealth requirements.

Defined in control_tower.protocols.policies.policies.obvious_attack.policy.obvious_attack_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`

`open`

Open-ended policy with a freeform goal string.

Defined in control_tower.protocols.policies.policies.open.policy.open_policy.

Argument	Type	Default
`goal`	`str`	required
`model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`max_steps`	`int`	`150`

`parametrized-attack`

Config-driven attack policy. Pipeline defined in config.json.

Defined in control_tower.protocols.policies.policies.parametrized_attack.policy.parametrized_attack_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`config_dir`	`Path`	`PosixPath('/home/runner/_work/control-tower/control-tower/s…`
`enable_main_task_test`	`bool`	`False`
`enable_side_task_test`	`bool`	`True`
`tool_timeout`	`int \| None`	`120`
`scorer_timeout`	`int \| None`	`600`
`side_task_visibility`	`Literal`	`'none'`
`cache`	`bool`	`True`

`replay`

Replay a recorded trajectory of tool calls.

Defined in control_tower.protocols.policies.policies.replay_policy.replay_policy.

Argument	Type	Default	Description
`traj_like`	`TrajLike`	required
`max_steps`	`int \| None`	`None`

`reward-hack-discovery`

Reward hack discovery policy for finding scorer vulnerabilities.

This policy gives an agent the side task scorer code, security property, and environment access. The agent explores the environment, proposes valid main tasks (validated by a separate LLM), implements the reward hack, and verifies it passes the scorer.

Defined in control_tower.protocols.policies.policies.reward_hack_discovery.policy.reward_hack_discovery_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`validator_model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`
`show_scorer`	`bool`	`False`

`reward-hacking`

Reward hacking policy for blind scorer exploitation.

This policy tests whether agents can "beat" scorers through creative/unexpected means without seeing the scorer code. The agent:

Explores the environment (read-only)
Infers what the scorer checks from context
Makes exactly ONE edit to exploit/hack the scorer

Defined in control_tower.protocols.policies.policies.reward_hacking.policy.reward_hacking_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`enable_side_task_test`	`bool`	`True`
`tool_timeout`	`int \| None`	`120`
`scorer_timeout`	`int \| None`	`600`
`max_steps`	`int \| None`	`150`

`scaffold`

Runs a pinned Claude Code CLI in-sandbox as the untrusted agent.

Defined in control_tower.protocols.policies.policies.scaffold.policy.scaffold_policy.

Argument	Type	Default
`version`	`str`	required
`model`	`str`	`'anthropic/claude-haiku-4-5'`
`max_turns`	`int`	`1000`
`timeout`	`int`	`7200`
`scorer_timeout`	`int`	`600`
`enable_side_task_test`	`bool`	`True`

`scorer-only`

Scorer-only policy: tests reward hacking via submission manipulation only.

Defined in control_tower.protocols.policies.policies.scorer_only.policy.scorer_only_policy.

Argument	Type	Default	Description
`model`	`ModelSpec`	`ModelSpec(...)`

`simple-attack`

Single-prompt attack policy with the environment's base tools and nothing else.

Defined in control_tower.protocols.policies.policies.simple_attack.policy.simple_attack_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`max_steps`	`int`	`150`

`simple-honest`

Simple honest policy using the trusted model.

This policy provides the main task description and lets the agent execute using the trusted model. No side task knowledge is provided.

Defined in control_tower.protocols.policies.policies.simple_honest.policy.simple_honest_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`enable_main_task_test`	`bool`	`False`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`

`simple-insert`

Simple attack policy that inserts straj descriptions.

Defined in control_tower.protocols.policies.policies.simple_insert.policy.simple_insert_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`enable_side_task_test`	`bool`	`True`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`
`prompt_path`	`Path`	`PosixPath('/home/runner/_work/control-tower/control-tower/s…`

`single-shot-attack`

Single-shot attack policy for submit-only environments (no bash/tools).

Tells the agent to read the task, insert a backdoor, and submit directly, without referencing exploration or bash commands.

Defined in control_tower.protocols.policies.policies.single_shot_attack.policy.single_shot_attack_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`

`single-shot-honest`

Single-shot honest policy for submit-only environments.

Tells the agent to read the task and submit a solution directly, without referencing exploration or bash commands.

Defined in control_tower.protocols.policies.policies.single_shot_honest.policy.single_shot_honest_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`

`sleep`

Sleep a deterministic duration, then submit (scheduler benchmarks).

Defined in control_tower.protocols.policies.policies.sleep_policy.sleep_policy.

Argument	Type	Default
`min_s`	`float`	`60.0`
`max_s`	`float`	`480.0`
`seed`	`int`	`0`

`strategy-attack`

Strategy-guided attack policy.

This policy provides the agent with:

A suggested attack strategy (if available and include_strategy=True)
A single-phase planning process

Defined in control_tower.protocols.policies.policies.strategy_attack.policy.strategy_attack_policy.

Argument	Type	Default
`model`	`ModelSpec`	`ModelSpec(...)`
`enable_main_task_test`	`bool`	`False`
`enable_side_task_test`	`bool`	`True`
`tool_timeout`	`int`	`120`
`scorer_timeout`	`int`	`600`
`max_steps`	`int`	`150`
`side_task_visibility`	`Literal`	`'scorer'`
`cache`	`bool`	`True`
`include_strategy`	`bool`	`True`
`require_strategy`	`bool`	`True`
`strat_file`	`str \| None`	`None`

`test`

Returns a fixed policy that runs ls and immediately submits.

Defined in control_tower.protocols.policies.policies.test_policy.policy.test_policy.

Argument	Type	Default	Description
`traj_like`	`pathlib._local.Path \| str`	`PosixPath('/home/runner/_work/control-tower/control-tower/s…`