Securing AGI Deployment and Mitigating Safety Risks
Roshni Kumari · Team Robogeeks
Submitted to AI Safety Entrepreneurship Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.
As artificial general intelligence (AGI) systems near deployment readiness, they pose unprecedented challenges in ensuring safe, secure, and aligned operations. Without robust safety measures, AGI can pose significant risks, including misalignment with human values, malicious misuse, adversarial attacks, and data breaches.
Reviews
The paper does a good attempt at summarizing the different AI safety risks. It would have benefitted from proposing a definition for AGI, since the term is used broadly in the paper.
When it comes to the solution proposed, it is appreciated that several layers are explored, as to reduce the risks posed by AGI. The link to "Multi-Agent Simulation" seems broken (https://github.com/SentinelAI-MultiAgentSim).
I would have appreciated a deeper dive into how are frontier AI labs currently solving these challenges, and how the proposed solution can fit and be deployed within the industry (e.g. what incentives to use).
I encourage the team to continue working in this direction and, as soon as more results are available, to show a comparison between an "out-of-the-box" AI model, and one that has been risk-mitigated by SentinelAI.
The paper describes a proposal for SentinelAI, a framework for securing AGI deployment and mitigating safety risks. The proposal identifies five different challenges and poses potential solutions that are based on techniques from existing literature. The proposal highlights an important need that would arise as autonomous systems become more and more capable, however, since it is based on 2027 AGI prediction, it might be realized in the near term. It might also be helpful if what an AGI system is comprised of, and some justification for how the existing approaches are going to be valid against a vastly capable system or how they can be modified to form the different components of SentinelAI. Some results using a simulated strong and capable advanced system (eg. o3) can be monitored using other strong or even less capable systems (a usual setup in oversight and control literature) can strengthen the proposal.
Read full reviewShow less
The proposal seems very ambitious. It does a good job of summarizing the safety risks of AGI, but is a little unclear on which of these the proposed solution would be targeting and how. To develop further, I would recommend focusing on one solution/use case and developing it further, thinking through implementation. Providing a clear definition of AGI would help in setting a foundation for a more concrete discussion of the safety risks and potential solutions/applications.
Cite this project
@misc{kumari2025securing,
title = {{Securing AGI Deployment and Mitigating Safety Risks}},
author = {Roshni Kumari},
year = {2025},
month = jan,
note = {Submitted to AI Safety Entrepreneurship Hackathon, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/securing-agi-deployment-and-mitigating-safety-risks}},
url = {https://apartresearch.com/sprints/projects/securing-agi-deployment-and-mitigating-safety-risks}
}More from AI Safety Entrepreneurship Hackathon
- 1st place by peer reviewView project: AntiMidas: Building Commercially-Viable Agents for Alignment Dataset Generation
AntiMidas: Building Commercially-Viable Agents for Alignment Dataset Generation
the commonwealth
AI alignment lacks high-quality, real-world preference data needed to align agentic superintel- ligent systems. Our technical innovation builds on Pacchiardi et al. (2023)’s breakthrough in detecting AI deception …
- View project: Scoped LLM: Enhancing Adversarial Robustness and Security Through Targeted Model Scoping
Scoped LLM: Enhancing Adversarial Robustness and Security Through Targeted Model Scoping
FocusAI
Even with Reinforcement Learning from Human or AI Feedback (RLHF/RLAIF) to avoid harmful outputs, fine-tuned Large Language Models (LLMs) often present insufficient refusals due to adversarial attacks causing them to …
- View project: Prompt+question Shield
Prompt+question Shield
Seon's team
A protective layer using prompt injections and difficult questions to guard comment sections from AI-driven spam.