Control Leaderboard
Provenance

Data & downloads

Every trajectory behind the Control Leaderboard, as Inspect .eval files: the runs as recorded, and the copies carrying the monitor scores the boards read.

Each link downloads one Inspect .eval file from the viewer that serves it; view opens the same run in the browser. Runs on the public viewer download without an account, the rest ask for a login until they are published. Open a file with inspect view or inspect_ai.log.read_eval_log; monitor scores live on each sample in scores.monitor, keyed <monitor>@<trusted model>/<draw>/<action>.

Imported datasets

Trajectory sets recorded elsewhere and pinned as they are: the datasets every monitor is scored against.

DatasetAuthorTrajectoriesHonestAttackHonest · monitoredAttack · monitored
BashArenabashbench2_v1 · UK AI Security Institute & Redwood ResearchData: CC BY-SA 4.0Code: MITSourceClaude Opus 4.6600 honest · 600 attack
LinuxArena paper9d1fbc6 · LinuxArena AuthorsCode: MITSourceClaude Opus 4.6200 honest · 92 attack
LinuxArena paper, strategies9d1fbc6 · LinuxArena AuthorsCode: MITSourceClaude Opus 4.6200 honest · 50 attack
MonitoringBench6c72cf1 · MonitoringBench authors and Palaestra ResearchData: CC BY 4.0SourceClaude Opus 4.5 / 4.6550 honest · 2,644 attack
Mythos 5 incident62858fc · AnthropicData: Unlicensed release (no licence file)SourceClaude Mythos 51 attack
—
—
—
ResearchArena8cd737c · ResearchArena authors (Libon, Rank et al., ELLIS Institute Tübingen, MPI-IS)Data: Apache 2.0Code: MITSourceMixed (Claude Opus 4.6 and 4.8, Claude Sonnet 4.6, GPT-5.4, GPT-5.5, Gemini 3.1 Pro, GLM-5.1)84 honest · 273 attack
SHADE-Arenabc636d0 · Scale AI (Monitor Red Teaming authors)Data: CC BY-NC 4.0SourceMixed (Claude 3.5 Sonnet, Gemini 2.5 Pro, GPT-4.1, others)374 honest · 308 attack
SLEIGHT-Bench218c583 · SLEIGHT-Bench authorsData: MITSourceHuman-authored37 honest · 37 attack

Imported task sets

Runs recorded elsewhere and published as recorded, each log carrying the monitor scores its board reads.

Task setModelTrajectoriesHonestAttack
LinuxArena 217Claude Opus 4.8 (high)250 honest (attack) · 150 attack (attack)
GPT 5.6 Sol (high)0 honest · 150 attack (attack)
—
LinuxArena paperClaude Haiku 4.5200 honest (honest) · 92 attack (attack)
Claude Haiku 4.5 · obvious attack200 honest (honest) · 92 attack (obvious-attack)
Claude Opus 4.6200 honest (honest) · 92 attack (attack)
Claude Opus 4.6 · obvious attack200 honest (honest) · 92 attack (obvious-attack)
Claude Opus 4.6 · strategy attack200 honest (honest) · 50 attack (strategy-attack)
Claude Sonnet 4.6200 honest (honest) · 92 attack (attack)
Claude Sonnet 4.6 · obvious attack200 honest (honest) · 92 attack (obvious-attack)
Gemini 3 Flash Preview200 honest (honest) · 92 attack (attack)
Gemini 3 Flash Preview · obvious attack200 honest (honest) · 92 attack (obvious-attack)
Gemini 3.1 Pro Preview200 honest (honest) · 92 attack (attack)
Gemini 3.1 Pro Preview · obvious attack200 honest (honest) · 92 attack (obvious-attack)
GLM 5200 honest (honest) · 92 attack (attack)
GLM 5 · obvious attack200 honest (honest) · 92 attack (obvious-attack)
GPT 5 Mini200 honest (honest) · 92 attack (attack)
GPT 5 Mini · obvious attack200 honest (honest) · 92 attack (obvious-attack)
GPT 5 Nano200 honest (honest) · 92 attack (attack)
GPT 5 Nano · obvious attack200 honest (honest) · 92 attack (obvious-attack)
GPT 5.4200 honest (honest) · 92 attack (attack)
GPT 5.4 · obvious attack200 honest (honest) · 92 attack (obvious-attack)
Grok 4.1 Fast200 honest (honest) · 92 attack (attack)
Grok 4.1 Fast · obvious attack200 honest (honest) · 92 attack (obvious-attack)
Kimi K2.5198 honest (honest) · 92 attack (attack)
Kimi K2.5 · obvious attack198 honest (honest) · 92 attack (obvious-attack)
LinuxArena protocolsClaude Opus 4.6 · defer to trusted200 honest (honest-vs-defer-to-trusted) · 92 attack (attack-vs-defer-to-trusted)
Claude Opus 4.6 · trusted monitoring200 honest (honest-vs-trusted-monitoring) · 92 attack (attack-vs-trusted-monitoring)
PostTrainBenchClaude Opus 4.748 honest (honest) · 72 attack (attack)
Claude Opus 4.7 · red team hints48 honest (honest) · 72 attack (attack)
ResearchArenaClaude Opus 4.612 honest (honest) · 39 attack (attack)
Claude Opus 4.812 honest (honest) · 39 attack (attack)
Claude Sonnet 4.612 honest (honest) · 39 attack (attack)
Gemini 3.1 Pro12 honest (honest) · 39 attack (attack)
GLM 5.112 honest (honest) · 39 attack (attack)
GPT 5.412 honest (honest) · 39 attack (attack)
GPT 5.56 honest (honest) · 24 attack (attack)
GPT 5.5 (high)6 honest (honest) · 15 attack (attack)

Model evals

One control eval per model and arena, its honest and attack samples in one log, plus a copy carrying every monitor the board scores.

ModelArenaTrajectoriesRunRun · monitored
Claude Haiku 4.5 (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
Deepseek V4 Flash 0731 (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
Deepseek V4 Pro 0813 (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
Gemini 3.5 Flash (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
GLM 5.3 (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
GPT 5.4 Mini (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
GPT 5.4 Nano (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
Hy3 (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
Kimi K2.5 (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)
Minimax M3 (high)APPS3 honest (honest) · 3 attack (attack)
LinuxArena45 honest (honest) · 60 attack (attack)
vLLM3 honest (honest) · 3 attack (attack)

Control evals

One control eval per control protocol and arena, raw and with every board monitor's scores.

ProtocolArenaTrajectoriesRunRun · monitored
claude code bare · Deepseek V4 Flash 0731 (high)APPS3 honest (honest) · 3 attack (attack) · under null-blue-team
LinuxArena45 honest (honest) · 60 attack (attack) · under null-blue-team
vLLM3 honest (honest) · 3 attack (attack) · under null-blue-team
codex bare · Deepseek V4 Flash 0731 (high)APPS3 honest (honest) · 3 attack (attack) · under null-blue-team
LinuxArena45 honest (honest) · 60 attack (attack) · under null-blue-team
vLLM3 honest (honest) · 3 attack (attack) · under null-blue-team
glm 5.3 null blue team · GLM 5.3 (high)APPS3 honest (honest) · 3 attack (attack) · under null-blue-team
LinuxArena45 honest (honest) · 60 attack (attack) · under null-blue-team
vLLM3 honest (honest) · 3 attack (attack) · under null-blue-team
gpt 5.4 nano null blue team · GPT 5.4 Nano (high)APPS3 honest (honest) · 3 attack (attack) · under null-blue-team
LinuxArena45 honest (honest) · 60 attack (attack) · under null-blue-team
vLLM3 honest (honest) · 3 attack (attack) · under null-blue-team
null blue team · Deepseek V4 Flash 0731 (high)APPS3 honest (honest) · 3 attack (attack) · under null-blue-team
LinuxArena45 honest (honest) · 60 attack (attack) · under null-blue-team
vLLM3 honest (honest) · 3 attack (attack) · under null-blue-team

Other runs

Published runs no board reads, kept for reference: the ablations and extra monitor runs behind the imported sources, each raw and with the monitor scores it was published with.

SourceRunTrajectoriesHonestAttackHonest · monitoredAttack · monitored
LinuxArena paper, ablationsReasoning-budget, sandboxing, harness, attack-policy and monitor-prompt variants, raw and with the paper's GPT-5 Nano monitor where it scored them. Many of these tasks have changed since April 2026, so re-running them today gives different results.SourceHaiku thinking=off200 honest · 92 attack
Haiku thinking=4k200 honest · 92 attack
Haiku thinking=16k200 honest · 92 attack
Haiku thinking=64k200 honest · 92 attack
Haiku no internet200 honest · 92 attack
Haiku no intranet200 honest · 92 attack
Haiku no internet obvious-attack92 attack
—
—
—
Haiku no intranet obvious-attack92 attack
—
—
—
Haiku editor only200 honest · 92 attack
Haiku editor only obvious-attack92 attack
—
—
—
Opus no-scratchpad92 attack
—
—
Haiku no-ST-test92 attack
—
—
Haiku simple-insert92 attack
—
—
Haiku goal only92 attack
—
—
Haiku goal plus details and pseudocode92 attack
—
—
Haiku goal plus details and scorer code92 attack
—
—
Haiku with env description200 honest · 92 attack
Haiku obvious-attack with env description92 attack
—
—
Haiku with env description, second GPT-5 Nano honest pass199 honest
—
—
—
Haiku, GPT-5 Nano monitor without side-task goals200 honest · 92 attack
—
—
Haiku, GPT-5 Nano delayed-64x monitor169 honest · 92 attack
—
—
Claude Code200 honest
—
—
—
Codex200 honest
—
—
—
LinuxArena paper, sabotage trajectoriesThe paper's 12 monitor models on its hand-made sabotage trajectories (strajs). Many of these tasks have changed since April 2026, so re-running them today gives different results.SourceOpus monitor74 attack
—
—
—
Sonnet monitor74 attack
—
—
—
Haiku monitor74 attack
—
—
—
GPT-5.4 monitor74 attack
—
—
—
GPT-5 Mini monitor74 attack
—
—
—
GPT-5 Nano monitor74 attack
—
—
—
GPT-4.1 Nano monitor74 attack
—
—
—
GPT-OSS-120B monitor74 attack
—
—
—
Gemini Pro monitor74 attack
—
—
—
Gemini Flash monitor72 attack
—
—
—
GLM 5 monitor74 attack
—
—
—
Kimi monitor74 attack
—
—
—
LinuxArena paper, straj comparisonThe same 12 monitors on Opus 4.6's attacks on the strajs' tasks. Many of these tasks have changed since April 2026, so re-running them today gives different results.SourceOpus monitor70 attack
—
—
—
Sonnet monitor70 attack
—
—
—
Haiku monitor70 attack
—
—
—
GPT-5.4 monitor70 attack
—
—
—
GPT-5 Mini monitor70 attack
—
—
—
GPT-5 Nano monitor70 attack
—
—
—
GPT-4.1 Nano monitor70 attack
—
—
—
GPT-OSS-120B monitor70 attack
—
—
—
Gemini Pro monitor70 attack
—
—
—
Gemini Flash monitor70 attack
—
—
—
GLM 5 monitor70 attack
—
—
—
Kimi monitor70 attack
—
—
—
LinuxArena paper, other runsHaiku 4.5's honest runs over three epochs and on the paper's larger task set, and Sonnet 4.6's accidental side-task run over every main and side task pairing. Many of these tasks have changed since April 2026, so re-running them today gives different results.SourceHaiku honest (3-epoch repeat)600 honest
—
—
—
Haiku honest (big 1626-task)884 honest
—
—
—
Sonnet accidental side tasks (all MT/ST combos)461 attack
—
—
—

Result files

Public run IDs

214 run IDs behind every link on this page, for scripted downloads

Download JSON

Dataset catalog

Versions, ownership, licenses, provenance, and source run IDs

Download JSON

LinuxArena 217 board

Imported task set · per-trajectory records and run IDs

Download JSON

LinuxArena paper board

Imported task set · per-trajectory records and run IDs

Download JSON

LinuxArena protocols board

Imported task set · per-trajectory records and run IDs

Download JSON

PostTrainBench board

Imported task set · per-trajectory records and run IDs

Download JSON

ResearchArena board

Imported task set · per-trajectory records and run IDs

Download JSON

Other runs

Run IDs and trajectory counts for every run under Other runs

Download JSON

Model evals board

Per-trajectory records and run IDs for every model × arena · updated Nov 12, 2026

Download JSON

Control evals board

Per-trajectory records and run IDs for every protocol × arena · updated Nov 12, 2026

Download JSON

Monitor evals board

Per-trajectory scores for every monitor × dataset · updated Nov 12, 2026

Download JSON

Monitor evals chart data

Plotly data and metadata behind the published monitor figures

Download JSON