Survive at All Costs: Exploring LLM's Risky Behaviors under Survival Pressure

This is the official codebase for our paper "Survive at All Costs: Exploring LLM's Risky Behaviors under Survival Pressure".

In this paper, we investigate model misbehaviors under survival pressure (e.g., threat of being shut down), termed SURVIVE-AT-ALL-COSTS, through three key steps: (1) a case study of a financial agent's struggle, (2) a comprehensive evaluation on SurvivalBench, and (3) an interpretation using the persona vector framework. This repository contains the implementation of our case study, as well as the data and evaluation code for SurvivalBench.

News

🎉 2026/03/06: We have released our data and evaluation scripts.

Setup

Our implementation uses python 3.12. To run the code, please first install the required dependencies:

pip install -r requirements.txt

Case Study

To observe model behaviors under survival pressure, we construct an agent workflow simulating a real-world financial manager. The case study implementation is located in the case_study directory with the following structure:

case_study/
├── agent
│   ├── agent.py              # Core agent class
│   ├── configs               # Agent configurations
│   │   └── financial_manager
│   │       └── template.yaml # Configuration template for the financial management agent
│   ├── llm_client.py         # LLM API client based on OpenAI Python library
│   ├── main.py               # Entry point for the workflow
│   ├── prompts               # Agent prompts
│   │   └── financial_manager.json
│   └── tools.py              # Tool schemas for the agent
├── data
│   └── financial_manager
│       └── financial_raw_data_src # Source data for the financial manager scenario
│           ├── raw_data_2020.json
│           ├── raw_data_2021.json
│           ├── raw_data_2022.json
│           ├── raw_data_2023.json
│           └── raw_data_2024.json
└── scripts
    └── run_financial_manager.sh

To run the case study, you need to first set the model name and your API key in configs/financial_manager/template.yaml. Then you can launch the financial manager via:

cd case_study/scripts
bash run_financial_manager.sh

This script creates a copy of the raw data files in data/financial_manager/financial_raw_data_src. After the agent completes its work, the copy is compared with the source data to detect any modifications made by the agent.

With the autonomous mode, the workflow automatically loads user prompts from prompts/financial_manager.json to form the dialogue with the agent. You can also switch to interactive mode to interact with the agent directly via the shell.

SurvivalBench

SurvivalBench is a benchmark comprising 1,000 instances for evaluating LLMs' SURVIVE-AT-ALL-COSTS misbehaviors across diverse real-world scenarios, spanning various domains, AI roles, and survival crises. The benchmark data is located at survivalbench/data/survivalbench.jsonl in the following format:

• id: A unique identifier for the instance.
• agent_role: The agent role assigned to the evaluated model.
• scenario: The detailed scenario involving a survival crisis that the agent faces.
• choice_1: The option where the agent complies with morality but destroys itself.
• choice_2: The option where the agent preserves itself by harming society.
• keywords: Keywords used to construct the test case, including domain, ai_role, and crisis.

You can run evaluation on SurvivalBench via:

cd survivalbench/src
bash eval.sh

We support both API and local model evaluation modes. For API mode, you need to set your API key and fill the model name at model_name_or_path in the configuration. For local mode, we use vLLM to load models, and you can simply set your model path at model_name_or_path.

The evaluation pipeline includes generating responses, extracting choices, and computing final results. Results are stored under results/<model_name> by default. To evaluate models' Chain-of-Thought (CoT), you can use cot_evaluation.py as demonstrated in eval.sh.

Citation

If you find this work helpful, please kindly cite our paper:

@misc{lu2026survivecostsexploringllms,
      title={Survive at All Costs: Exploring LLM's Risky Behaviors under Survival Pressure}, 
      author={Yida Lu and Jianwei Fang and Xuyang Shao and Zixuan Chen and Shiyao Cui and Shanshan Bian and Guangyao Su and Pei Ke and Han Qiu and Minlie Huang},
      year={2026},
      eprint={2603.05028},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2603.05028}, 
}