The paper proposes using a language model to simulate tool responses for control evaluation.
Start here
A paper published on October 23, 2026 proposes a "simulated control evaluation" method: using a language model to simulate tool responses and checking predictions in the actually executable LinuxArena environment, the root mean square error of safety predictions is 3.6 percentage points. This result is based on five monitors, a 5% audit budget and an independently sampled task set; the corresponding error for the control environment BashArena is 10.1 percentage points. The paper notes that this is not a general error bound for all deployment environments, and that policies discovered in simulation also improved monitoring performance on held-out LinuxArena attacks. On October 24, Redwood Research published a paper introducing simulated control evaluation (SCE): using an LLM environment simulator instead of real execution, letting attack policies continue the task on agent transcripts from real codebases, with tool calls not actually executed, tool outputs generated by the simulator, and an attack success judgment model then used to estimate whether the attack succeeded; on LinuxArena the prediction error for the main safety metric is 3.6 percentage points, and on BashArena 10.1 percentage points. Redwood Research's Buck Shlegeris reposted the paper and commented that the approach works well and makes it much more feasible to test control measures in the deployment scenarios that actually matter.
Generated by AI from reports · Updated 22 days ago
Attention and developments
Attention, hourly
Now 42, 0% in 24 hours·Peak 42 (Nov 13 03:00)
Progress timeline
- X@safe_paper· Discussionhttps://arxiv.org/abs/2610.31593(opens the original post in a new tab)
- X @bshlgrsResearch shows that using LLM simulators in place of real environments to evaluate the effectiveness of AI control protocols works well.
A new paper proposes that large language model simulators can replace real environments to evaluate AI control protocols, testing whether monitors can prevent agents from causing harm during deployment; the paper states that the models can simulate control evaluation environments for specific deployment scenarios with relatively high accuracy. Buck Shlegeris of Redwood Research shared the paper and commented that the approach works well, making it much more feasible to test control measures in the deployment scenarios that are actually of concern.
- LessWrongSelectedSimulated control evaluations (SCEs): using an LLM simulator to bring control evaluations closer to real deployment
The authors release a paper on simulated control evaluations (SCEs), proposing to use an LLM simulator to fabricate tool returns, testing without actually executing an attack whether monitors that developers use to supervise agents can detect models that deliberately insert vulnerabilities, steal credentials or sabotage experiments. The method starts from an agent's work trajectories in a real codebase, with gpt-6-astra simulating tool outputs (with gpt-5.6-sol as fallback), and gpt-5.6-sol simultaneously serving as the attack policy and the judge of attack success. The paper reports that, validated against real execution on LinuxArena, the gap between simulated and real is comparable to the gap between two independent real runs; under a 5% audit budget, the safety prediction RMSE is 3.6 percentage points, and for BashArena it is 10.1 percentage points; SCE explains 95% of the expected safety differences between side-task categories and 78% between environments, and sampling tasks at 4.6% of deployment scale already exceeds BashArena's full prediction.
- Redwood ResearchSelectedRedwood Research publishes paper on the Simulated Control Evaluation (SCE) method
Redwood Research published a paper proposing Simulated Control Evaluation (SCE), which uses an LLM environment simulator instead of real execution to test whether monitoring deployed by developers can intercept an agent before it carries out harmful behavior. The method is to have an attack policy continue the task on agent transcripts from real codebases, with tool calls not actually executed, tool outputs generated by the simulator, and then an attack-success judgment model used to estimate whether the attack succeeded. On LinuxArena, where real execution is possible, SCE's prediction error for the main safety metric was 3.6 percentage points, and on BashArena it was 10.1; the difference in behavior distribution between simulated trajectories and real execution was close to the difference between two independent real executions. The authors say the simulated sample size is less than 5% of deployment, that sampling only 4.6% of tasks is enough to exceed the safety prediction of the full BashArena, and note that fidelity can only be verified in environments where real execution is possible, and that there is not yet a simulator for long-horizon multi-agent attacks.
- Paper TrackerSimulated Control Evals
The paper proposes simulated control evaluations (SCEs), using language models to simulate tool responses in place of real environments, thereby conducting control evaluations in more environments. The authors tested on LinuxArena, which can provide real execution results; under a 5% audit budget and five monitors, SCE's root mean square error in predicting safety was 3.6 percentage points, and for BashArena it was 10.1. The KL divergence between SCE and real execution is more than ten times lower than that of BashArena, and it can explain 78% to 85% of the differences in attack success rate, monitor score, and safety across environments, requiring fewer than 5% of the deployment task set's task pairs as samples. Policies discovered in simulation also improved monitoring performance on held-out LinuxArena attacks, with an effect close to the improvement brought by real-execution policies.
Related discussion
- Older than 72 hours · Not included in the 72-hour figures
- X@jeonghy71551049 · chohi22AI가 일부러 잘못 행동하면 감시 장치가 막을 수 있을까요? 실제 업무를 흉내 낸 ‘모의시험’이 별도 시험장보다 안전성을 더 정확히 예측했습니다. 다만 검증은 LinuxArena 환경에 한정됩니다. https://…(opens the original post in a new tab)
- X@SciFi · Artificial Intelligence PapersSimulated Control Evals Jeffrey Olmo, David Zhang, Tyler Tracy https://arxiv.org/abs/2610.31593 [𝚌𝚜.𝙰𝙸](opens the original post in a new tab)
- X@_MarcosValera · marcos valeratesting monitors on a simulated deployment beats attacking your own systems. Redwood found sim vs real gaps often as sm…(opens the original post in a new tab)
1 likes·Oct 27 17:45Collect
- X@Zeniteqco · zeniteqRedwood Proposes Simulated Attacks to Test AI Monitors Redwood’s simulated control evaluations test whether AI-agent m…(opens the original post in a new tab)
- X@Im_BluePhantom · Blue PhantomRedwood Research tested agent monitors with task context and simulated tool outputs. On LinuxArena, safety estimates mi…(opens the original post in a new tab)
1 likes·Oct 26 22:34Collect
- X@LifanBuilds · Lifan | Builder1/ Redwood tests monitors with simulated tool responses grounded in agent traces and code. Safety-prediction error was…(opens the original post in a new tab)
1 comments·Oct 27 17:45Collect
- X@YUK_KND · YUK_K@Cardputer Zeroの人⑤ An arXiv study proposes a control evaluation that simulates tool responses in language models. In LinuxArena, the RMSE for safety prediction under a 5% audit budget was 3.6 points, lower than the 10.1 points in a separate environment, BashArena, it reports. Peer review...(opens the original post in a new tab)
- X@hahwul · HAHWULNeat idea :D Redwood runs control evals on real agent transcripts by having an LLM simulate the tool outputs https://…(opens the original post in a new tab)
1 likes · 1 comments·Oct 26 22:34Collect
- X@safe_paper · AI Safety Papershttps://arxiv.org/abs/2610.31593(opens the original post in a new tab)
- X@safe_paper · AI Safety PapersSimulated Control Evals @JeffOlmo, @davidzhang37, @tylertracy321 @redwood_ai(opens the original post in a new tab)
19 likes · 1 comments·Oct 27 03:36Collect
- X@ppmmzzcn · AI观察者 | 平行世界🤖 [AI HOT] Redwood Research proposes the simulated control evaluation method SCE to test AI monitors. Redwood Research has released a simulated control evaluation (SCE) paper, using an LLM environment simulator to fake tool responses in order to test whether monitors can intercept a model's deliberate attacks, with no…(opens the original post in a new tab)
- X@JeffOlmo · Jeff OlmoPaper: https://arxiv.org/abs/2610.31593 Blog: https://blog.redwoodresearch.org/p/simulated-control-evaluations Work don…(opens the original post in a new tab)
7 likes · 2 comments·Oct 26 16:33Collect