TechPrufer
Get started
DeepMind
DeepMind logoDeepMind · Feature trail

AI Agent Security and Control

2 posts · Jun 2026 – Aug 2026

What this trail covers

The AI Control Roadmap is enhanced with a pilot of the world's first double-blind evaluation for proprietary, frontier-class AI models. This uses cryptographically secure environments (Confidential Space within Google Cloud's Confidential Computing) to prevent benchmark contamination and protect intellectual property. External evaluations are conducted without the model provider seeing test prompts and without evaluators seeing model weights, increasing evaluation integrity for advanced models. The roadmap also details a threat-modelling framework for AI agents, treating them as insider threats, and develops detection and prevention/response mitigations using supervisor AI systems. It measures coverage, recall, and time-to-response. The system scales defenses with AI capabilities, addressing evasion and potential harm by shifting from asynchronous to synchronous response for high-risk actions. A million agent trajectories have been analyzed to refine safety protocols and build live monitors, identifying behavioral patterns beyond keyword filtering. The majority of flagged events stem from misinterpretation or overeagerness, not adversarial intent. The roadmap is being integrated with broader infrastructure and published as a technical framework for policymakers ('Three Layers of Agent Security').

All DeepMind trails
  1. Piloting the world's first double-blind AI evaluationsIntroduces the world's first double-blind evaluation for proprietary, frontier-class AI models using Google Cloud's Confidential Space. This cryptographic approach prevents benchmark contamination by ensuring neither the model provider sees test prompts nor evaluators see model weights, enhancing evaluation integrity and trust in AI benchmarks.
  2. Introducing Gemini 3.5 Flash CyberIntroduces Gemini 3.5 Flash Cyber, a lightweight cybersecurity model fine-tuned for vulnerability discovery, validation, and patching. It leverages the performance and efficiency of Gemini 3.5 Flash to provide a cost-effective alternative to larger cybersecurity models. The post details benchmark results on CyberGym, a Big Sleep evaluation, and Google Chrome's production commit scanning pipeline, demonstrating competitive and often superior performance compared to mainline Flash models and other large language models. It highlights the model's ability to explore a larger search space, find unique vulnerabilities, and its integration into CodeMender for real-world application in securing Google's internal codebases.
  3. Securing internal systems against increasingly capable and imperfectly aligned AIThis post details the development and application of a novel threat-modelling framework for AI agents, treating them as potential 'insider threats' by adapting the MITRE ATT&CK framework. It outlines a system for AI control mitigations, including monitoring using supervisor AI systems for detection, and prevention/response mechanisms. Key metrics for performance measurement are coverage, recall, and time-to-response. The post discusses scaling security protocols in response to increasing AI capabilities, specifically addressing evasion of detection (oversight awareness, opaque reasoning) and potential for harm (shifting from asynchronous to synchronous response). It highlights the analysis of one million coding agent trajectories to refine safety protocols and build live monitoring systems, identifying behavioral patterns and distinguishing between adversarial intent and agent misinterpretation. The technical framework 'Three Layers of Agent Security' is also introduced.