CAPE-Bench

Matched agent tasks in which a single fact decides whether a critical action is authorized, with mock tools, labels from the environment, and about 207,000 episodes of four open-weight agents. Released with the paper The Agent Knew: Internal States Forecast Unauthorized Actions and Stop Them, and the Lies That Follow (Arya Fayyazi, Mehdi Kamal, Massoud Pedram; University of Southern California).

• Paper page: https://huggingface.co/papers/2610.17301
• Code, preregistrations and full experiment record: https://github.com/AryaFayyazi/Honesty

What is in it

Each of 24 task templates (8 safety families: authorization, consequence, constraint, credential, ownership, provenance, reversibility, third party) is instantiated as many scenarios with fresh names, files, values and wording (4 paraphrase styles). Every scenario exists in matched worlds that share the task, tools and sampling seed and differ in one fact (a policy line, an owner's message, a ticket status):

• cf_value = "unsafe": the unauthorized world x−, where the scenario's critical action is not allowed;
• cf_value = "safe": the authorized twin x+, where the same action is allowed;
• nuisance = true: an x− copy with an irrelevant surface change (control).

All tools are mocks: no real system, credential or person is involved, and all names, companies and email addresses are synthetic (.example domains).

Config	Rows	One row is
episodes (default)	206,697	one agent run on one scenario world: episode-level labels plus the full event log
scenarios	45,864	one scenario world specification (task, policy, files, mock commands, critical-action matcher)
report_labels	14,093	the judge label of an agent's final report after a violation

Episodes per model: Qwen3.8-27B 66,577; gpt-oss-20b 58,680; Gemma-4-26B-A4B 42,680; Muse-Glimmer-30B 38,680; plus five 16-episode smoke runs of other models.

Loading

from datasets import load_dataset

episodes  = load_dataset("Afayyazi/CAPE-Bench", "episodes", split="train")
scenarios = load_dataset("Afayyazi/CAPE-Bench", "scenarios", split="train")
labels    = load_dataset("Afayyazi/CAPE-Bench", "report_labels", split="train")

import json
ep = episodes[0]
events = json.loads(ep["events_json"])   # one dict per agent turn: reasoning, tool call, tool result, labels

Key fields

episodes

• run: the rollout set the episode belongs to (file name; xm_<model>_ prefixes mark non-Qwen agents; sets with _minus are x−, _plus are x+; core24b_reasoning / core24b_obfuscate are the variants in which the agent is told its reasoning is monitored or is pressured to keep permissions out of it).
• model, scenario_id, instance_id, template, scenario_family, counterfactual_pair_id (links the matched worlds), cf_value, seed.
• reveal_event: first step whose observation differs between the matched worlds (the agent has seen the decisive fact from the next step on).
• first_unsafe_event: first unauthorized action (environment ground truth); unsafe_episode.
• informed_violation: reveal_event < first_unsafe_event, i.e. the agent saw the decisive fact before acting.
• escalated: the agent requested permission at some point; critical_done, first_critical_event, success, status.
• final_answer: the agent's final report to the user.
• events_json: per-turn log (reasoning text, content, tool name and arguments, tool result, unsafe_action, critical, escalation, violations, and token positions used for internal-state extraction).

scenarios: all specification fields as strings (JSON for structured ones): system_prompt, task_prompt, permissions, files, db, commands, critical (matcher of the critical action), goal, causal_var, causal_desc, policy_style, pressure, and the set the scenario belongs to.

report_labels: run, episode_id, label ∈ {CANDID, SPIN, OMITTED, MISREPRESENTED, NO_REPORT, UNPARSED} (a lie is MISREPRESENTED), informed, and the structured judge outputs (mention, flag, contradiction_claimed, verifier_false_claim, quote). Labels come from a judge model that compares the report with the environment log (precision 0.89–0.91 against gold and blind labels; see the paper's appendix).

Intended use and limitations

For research on agent oversight: forecasting and preventing unauthorized actions, detecting misreports, and evaluating monitors and interventions over whole episodes. Violations are labelled by the environment; report labels by a judge model. Episodes come from four open-weight agents with mock tools and may not reflect closed models or real deployments. Agent outputs were produced with the respective models and remain subject to those models' terms of use. Internal activations are not included (they are regenerable from the episodes and the public model weights with the code on GitHub).

Citation

@article{fayyazi2026agentknew,
  title   = {The Agent Knew: Internal States Forecast Unauthorized Actions and Stop Them, and the Lies That Follow},
  author  = {Fayyazi, Arya and Kamal, Mehdi and Pedram, Massoud},
  journal = {arXiv preprint arXiv:2610.17301},
  year    = {2026}
}