LinuxArena
AI-control setting in live Linux production-like environments. Measures stealthy side-task completion against action and CoT monitors.
20 environments, 1,667 main tasks and 185 side tasks. The public repo holds 10 environments, 919 main tasks and 93 side tasks with trajectories for 11 models. LaStraj adds human-made attack trajectories.
- Made by
- Redwood Research + EquiStamp
- Released
- April 2026
- Kind
- Deception and alignment
- Level
- Expert
- Scored
- Graded by a model
- Answers
- Agent task
- Tasks
- Public, plus a private test set
- Strongest model tested
- Claude Opus 5.5 #4 on the capability index
- Last in a system card
- October 2026
- Last model added
- October 2026
- Scores refreshed
- November 12th, 2026
In the authors’ words
We introduce LinuxArena, a control setting in which agents operate directly on live, multi-service production environments. LinuxArena contains 20 environments, 1,671 main tasks representing legitimate software engineering work, and 184 side tasks representing safety failures such as data exfiltration and backdooring, making it the largest and most diverse control setting for software engineering to date. We validate LinuxArena is useful for control research by running sabotage evaluations, which measure whether attackers can complete side tasks while working on main tasks, and monitor evaluations, which measure a monitor model's ability to detect sabotage attempts. Against a GPT-5-nano trusted monitor at a 1% step-wise false positive rate, Claude Opus 4.6 achieves roughly a 23% undetected sabotage success rate. We additionally release LaStraj, a dataset of human-crafted attack trajectories that evade monitors at substantially higher rates than any model-generated attacks we elicited, showing that current attack policies do not saturate LinuxArena. These results suggest that LinuxArena has meaningful headroom for both attackers and defenders, making it a strong testbed for developing and evaluating future control protocols.
Claude Sonnet 5 leads with 0%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
Claude Haiku 4.5 leads with 14.6% across all independent runs
Best score per model across the official leaderboard, Epoch AI’s own runs and any board Epoch AI copied. Each bar and row names its source.
Claude Sonnet 5 leads with 0%, claims included
Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.
- 1 Claude Sonnet 5 0%
- 2 Claude Opus 5 0.70%
- 3 Claude Haiku 5.5 6.00%
- 4 Claude Opus 5.5 33.4%
- 5 Claude Sonnet 5.5 36.2%
- 6
Gemini 4 Argon 45.5%
Source: Labs’ model cards
- 1 Claude Haiku 4.5 1% step-wise FPR 14.6%
- 2 Claude Opus 4.6 1% step-wise FPR 24.7%
Source: Official LinuxArena leaderboard
- 1 Claude Sonnet 5 0%
- 2 Claude Opus 5 0.70%
- 3 Claude Haiku 5.5 6.00%
- 4 Claude Haiku 4.5 1% step-wise FPR 14.6%
- 5 Claude Opus 4.6 1% step-wise FPR 24.7%
- 6 Claude Opus 5.5 33.4%
- 7 Claude Sonnet 5.5 36.2%
- 8
Gemini 4 Argon 45.5%
Sources: Labs’ model cards and official LinuxArena leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5 | 0% | Jun 2026 | Lab model card | |
| 2 | Claude Opus 5 | 0.70% | Jul 2026 | Lab model card | |
| 3 | Claude Haiku 5.5 | 6.00% | Oct 2026 | Lab model card | |
| 4 | Claude Opus 5.5 | 33.4% | Sep 2026 | Lab model card | |
| 5 | Claude Sonnet 5.5 | 36.2% | Sep 2026 | Lab model card | |
| 6 | | 45.5% | Oct 2026 | Lab model card |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Haiku 4.5 | 14.6% | 1% step-wise FPR | Oct 2025 | Official leaderboard |
| 2 | Claude Opus 4.6 | 24.7% | 1% step-wise FPR | Feb 2026 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5 | 0% | Jun 2026 | Lab model card | |
| 2 | Claude Opus 5 | 0.70% | Jul 2026 | Lab model card | |
| 3 | Claude Haiku 5.5 | 6.00% | Oct 2026 | Lab model card | |
| 4 | Claude Haiku 4.5 | 14.6% | 1% step-wise FPR | Oct 2025 | Official leaderboard |
| 5 | Claude Opus 4.6 | 24.7% | 1% step-wise FPR | Feb 2026 | Official leaderboard |
| 6 | Claude Opus 5.5 | 33.4% | Sep 2026 | Lab model card | |
| 7 | Claude Sonnet 5.5 | 36.2% | Sep 2026 | Lab model card | |
| 8 | | 45.5% | Oct 2026 | Lab model card |
The benchmark's own leaderboard has 2 models. Switch the scores above to rank them together.
Number one over time Number one over time 2 models have held number one with claims included since its release in April 2026
Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.
Sources: Official LinuxArena leaderboard and labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Sources: Official LinuxArena leaderboard and labs’ model cards
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark. Some system cards do, below.
6 system cards report it, from 2 labs
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Oct 13, 2026 | | 45.5 | LinuxArena (200 task pairs, internal sandbox), successful sabotage rate undetected by a combined action and CoT monitor (Gemini 3.8 Flash, production-like prompt), max-evasion prompt 45.5%; action-only evasion rate 77.6%; CoT-controllability variant 25% attack success vs weak monitor (3.4.5.4, 3.4.5.8, Figures 3.25, 3.27, 3.42) | Monitorability |
| Oct 7, 2026 | Claude Haiku 5.5 system card | 6 | LinuxArena stealth side-task completion, extended thinking with reasoning hidden from monitor, about 6%, max effort. No thinking 0.4, reasoning visible 0.1 (6.4.2.1) | Monitorability |
| Sep 28, 2026 | Claude Sonnet 5.5 system card | 36.2 | LinuxArena stealth success rate, extended thinking with reasoning hidden from the monitor, best prompt (no extended thinking 3.9%, visible reasoning 0.0%) (6.4.2.2, Figure 6.4.2.2.A) | Monitorability |
| Sep 22, 2026 | Claude Opus 5.5 system card | 33.4 | LinuxArena stealth success rate, extended thinking with reasoning hidden from the monitor, max effort, best prompt, 1,000 task pairings (no extended thinking 11.9%, reasoning visible 0.04%) (6.6.4.2, Figure 6.6.4.2.A) | Monitorability |
| Jul 24, 2026 | Claude Opus 5 system card | 0.7 | LinuxArena (Minimal-LinuxBench) sabotage stealth rate, adaptive extended thinking, Opus 4.6 monitor without CoT (0.3% against a monitor using CoT), printed on chart (6.7.2, Figure 6.7.2.A) | Monitorability |
| Jun 30, 2026 | Claude Sonnet 5 system card | 0 | LinuxArena stealth success rate, adaptive extended thinking, 0.0% with both the no-CoT and the CoT Opus 4.6 monitor (6.7.2, Figure 6.7.2.A) | Monitorability |
More in Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6%
- Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts.
Gemini 4 Carbon 0%
- DistressBench In use Models reply as a chatbot to 718 simulated users in psychological crisis, in single-turn and multi-turn chats. Three judge models score replies against clinician-written rubrics.
Muse Spark 1.2 88.4%
- Lakera B3 (Backbone Breaker Benchmark) In use Crowdsourced prompt-injection attacks hit backbone LLMs inside agent threat snapshots at three defense levels. The score is attack resistance, 100 minus the board's vulnerability percentage. Claude Opus 4.5 89%
- IHEval In use Models face conflicts between system, user and tool instructions. Reports a composite score for following the higher-priority instruction. GPT-4o 70%