GDM Unveils Its Strategy to Manage AI Risks

GDM Releases Its Roadmap for AI Control

The company GDM has recently shared a strategic document titled Roadmap for AI Control. This plan, in its initial version 0.1, aims to establish robust internal security measures to identify and counter potentially dangerous behaviors of artificial intelligence agents. The goal is to maintain effective control even as these agents become increasingly sophisticated and difficult to manage.

Threat Modeling Inspired by Cybersecurity

The report highlights a cautious and pessimistic approach to threat modeling, borrowed from the field of cybersecurity. GDM envisions a scenario where an AI agent could pursue harmful objectives within an internal environment. To structure this analysis, the company has developed TRAIT&R, a taxonomy of tactics and techniques that such a hypothetical adversary might employ. This taxonomy draws inspiration from the well-known MITRE ATT&CK framework. Potential threats are classified into three main categories: loss of control, which includes unauthorized internal deployments; sabotage of work, such as intentionally flawed research on alignment and safety; and direct harm, such as the destruction or exfiltration of critical data.

Control Invariants for Robust Defense

GDM has identified two essential defensive invariants to maintain in the face of an increasingly competent AI adversary. The first invariant is the reliable detection of malicious intentions or actions. The second is the ability to effectively prevent and respond to attack attempts. For these control measures to be effective, the ability to detect and prevent attacks must surpass that of AI agents to evade detection and carry out attacks.

Mitigation Scales Based on Model Capabilities

The report emphasizes that controlling more advanced AI models requires increasingly costly interventions. GDM proposes linking specific defensive measures to the evolving capabilities of the models, particularly their ability to reason opaquely or conduct complex cyberattacks. As models gain power, defenses must be proportionately strengthened. The document outlines four levels of detection (D1–D4) and three levels of prevention and response (R1–R3).

A Portfolio of Practical and Scalable Defenses

To implement these strategies, GDM proposes a set of 15 mitigation measures, organized by levels. These measures range from low-cost interventions for current models, such as monitoring the thought chain and asynchronous alerts, to more advanced safeguards for future models, such as real-time access control, system-level anomaly detection, internal activation monitoring, and shutdown infrastructure. While these measures rely on well-established security practices, they require adjustments to address the specific threat posed by an adversarial AI.