Collusion and Mitigation in AI Control
VAXimum Security
Submitted to AI Control Hackathon 2025. Sprint projects are early-stage work by participants, not Apart Research publications.
As large language models (LLMs) grow more powerful and operate with greater autonomy, ensuring their safety becomes increasingly critical. While prior research has introduced safety protocols and evaluated their resilience against adversarial model behavior, a critical gap remains: existing evaluations have not addressed whether these techniques remain effective when the system’s monitor is compromised or colludes with the model. In this hackathon, we designed and demonstrated an effective attack that targets and compromises the monitor. To address this vulnerability, we introduce new safety protocol pipelines—conceptually inspired by vaccines—that safeguard both the monitor and the executor.
Reviews
No public critique yet.
Cite this project
@misc{security2025collusion,
title = {{Collusion and Mitigation in AI Control}},
author = {VAXimum Security},
year = {2025},
month = mar,
note = {Submitted to AI Control Hackathon 2025, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/collusion-and-mitigation-in-ai-control-pm60}},
url = {https://apartresearch.com/sprints/projects/collusion-and-mitigation-in-ai-control-pm60}
}More from AI Control Hackathon 2025
- 1st place by peer reviewView project: Can models use their Chain-of-Thought to attack overseers?
Can models use their Chain-of-Thought to attack overseers?
This project explores vulnerabilities in AI evaluation mechanisms, specifically focusing on how AI agents might influence their overseeing AI judges through manipulative instructions embedded within their …
- 2nd place by peer reviewView project: Schelling coordination via agentic loops
Schelling coordination via agentic loops
Focal
Capability to collude undermines untrusted monitoring protocols. We propose a novel way to evaluate acausal coordination capabilities of LLMs via repeated sampling in causally separated agentic loops. We provide an …
- 3rd place by peer reviewView project: Model Models: Simulating a Trusted Monitor
Model Models: Simulating a Trusted Monitor
Charlie's Angels
We offer initial investigations into whether the untrusted model can 'simulate' the trusted monitor: is U able to successfully guess what suspicion score T will assign in the APPS setting? We also offer a clean, modular …