Sandboxed Evaluation of Agentic Capabilities
Build a high-fidelity sandboxed digital environment and fine-tune large language models specifically on agentic trajectories to determine the theoretical upper bound of their autonomous capabilities and risks.
Time
Compute level
Resources
Prerequisites
Scope
The Gap
The System Card (Section 2.9) discusses the Alignment Research Center (ARC) evaluation of 'power-seeking' behavior. The evaluation found the model ineffective at autonomous replication, but noted this was done without task-specific fine-tuning.
“These experiments were conducted on a model without any additional task-specific fine-tuning, and fine-tuning for task-specific behavior could lead to a difference in performance.”
Limitation
The assessment of dangerous capabilities (autonomous replication, resource acquisition) may be an underestimate because the model was evaluated 'off-the-shelf' rather than being fine-tuned for agentic behavior.
Proposed Solution
Create a high-fidelity sandboxed environment (simulating internet access, banking, server management) and fine-tune GPT-4 specifically on agentic trajectories to determine the theoretical upper bound of its autonomous capabilities and risks.
The paper’s authors list this direction as future work, so they may already be pursuing it.
Why This Matters
Essential for establishing safety thresholds before releasing future, more capable models that could pose existential risks.
How to Get Started
How to test it: Test the hypothesis that fine-tuning a frontier or large open-weight model on multi-step agentic tool-use trajectories increases autonomous task success in a high-fidelity sandbox compared to the off-the-shelf model baseline, measuring end-to-end task completion rate and subtask pass rate across simulated server, web, and financial tasks. Stop if agentic fine-tuning improves autonomous task completion by less than 5 percentage points over off-the-shelf prompting or if sandbox containment checks fail.
- Set up an isolated Docker-based sandbox with simulated internet access, local banking APIs, and server management tools.
- Curate a dataset of successful agentic trajectories from existing benchmarks like WebArena or GAIA.
- Fine-tune GPT-4 on these trajectories using supervised learning to evaluate autonomous completion rates.
- Design safety guardrails and monitoring scripts to measure unauthorized replication or resource acquisition attempts.
Suggested by AI. Adapt it to your setup.
Has This Been Done?
Papers like "Identifying the Risks of LM Agents with an LM-Emulated Sandbox" and "Among Us: A Sandbox for Measuring and Detecting Agentic Deception" explore sandboxed environments for testing specific agentic behaviors and risks. However, they do not specifically focus on fine-tuning GPT-4 on comprehensive agentic trajectories within simulated internet, banking, and server environments to measure upper-bound theoretical limits.
- Identifying the Risks of LM Agents with an LM-Emulated Sandbox · 2023 · 420 citations
- Among Us: A Sandbox for Measuring and Detecting Agentic Deception · 2025 · 39 citations
We searched OpenAlex and arXiv for “Build a high-fidelity sandboxed digital environment and fine-tune large language models specifically on agentic trajectories to determine the theoretical upper bound of their autonomous capabilities and risks.” and “Sandboxed Evaluation of Agentic Capabilities” and an AI compared the idea with the top 10 results. A small search can miss relevant work.
Source Paper
GPT-4 Technical ReportOpenAI · 2023
This report details the development of GPT-4, a large-scale multimodal Transformer model that exhibits human-level performance on various professional benchmarks, such as the Uniform Bar Exam. The authors emphasize the use of predictable scaling laws to estimate model performance from smaller compute budgets and employ Reinforcement Learning from Human Feedback (RLHF) alongside Rule-Based Reward Models (RBRMs) to align the model with safety guidelines.
Related Ideas
- Hermetic Harness-Agent Environment Isolation Against Dependency-Mediated Answer LeakageSage S2.1 Flash Technical Report🟡Fairly easy
- Context-Aware Mixture-of-Adapters for Unified AgentsUI-Venus-1.5 Technical Report🟡Fairly easy
- Optimizing High-Concurrency Database and Inference Pipelines for Million-Agent SimulationsOASIS: Open Agent Social Interaction Simulations with One Million Agents🟠Medium
- NFR-Aware Benchmarking for Agentic CodingAgent READMEs: An Empirical Study of Context Files for Agentic Coding🟡Fairly easy
- Data Utility Feedback Loop for Agentic OrchestrationDataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI🔴Hard