Towards Cybersecurity SuperIntelligence (CSI): What's the best harness for cybersecurity?
What is the best agent harness for cybersecurity? Benchmarking five heterogeneous scaffolds (CSI::Claude, ::Codex, ::Mistral, ::GCAI, ::CAI) on 33 Cybench challenges with a fixed model (alias2-mini), no single scaffold dominates — each contributes a unique solve the others miss. A blackboard multi-agent architecture that lets scaffolds exchange typed findings reaches 19/33 (57.6%), a 27% relative gain over the best individual scaffold (15/33, 45.5%) — 25% faster at comparable cost. Heterogeneity of scaffolds is the structural lever toward cybersecurity superintelligence.
Cybersecurity AI (CAI) Dataset
A 14-month corpus of cybersecurity LLM trajectories collected through the open-source CAI framework: 230,935 session logs, 26 M user prompts from 16,768 source IPs across 123 countries, totalling 18.07 TB of durable storage. Hands-on by construction (36% offensive, 20% attacker-intent, 28% business, 4% defensive) — the largest described corpus of LLM-driven hacker trajectories. It exposes how operators routinely paste live credentials, hostnames and tokens into prompts they know are being logged, motivating the case for on-premise, operator-trust-boundary cybersecurity LLMs.
Dynamic Cyber Ranges
LLM-powered cyberattack agents are rapidly outperforming traditional benchmarks like Jeopardy-style CTFs and static cyber ranges, exposing the need for more adaptive evaluation environments. Dynamic Cyber Ranges—augmented with AI-driven defender agents—significantly reduce attack success (down to 0–55%) and maintain robustness as models evolve, with even smaller on-premise models proving highly effective at defense.
Cybersecurity AI: Hacking Consumer Robots in the AI Era
Has generative AI compromised robot cybersecurity? What historically required deep knowledge of ROS, ROS 2, and robotic system internals can now be automated by anyone using AI tools. Three case studies demonstrate vulnerabilities in consumer robots: an autonomous lawnmower, a powered exoskeleton, and a window-cleaning robot, with CAI discovering 38 vulnerabilities automatically.
Towards Cybersecurity Superintelligence: from AI-guided humans to human-guided AI
Cybersecurity superintelligence -- artificial intelligence exceeding the best human capability in both speed and strategic reasoning -- represents the next frontier in security. This paper documents the emergence of such capability through three major contributions that have pioneered the field of AI Security.
Cybersecurity AI: A Game-Theoretic AI for Guiding Attack and Defense
Can we make Cybersecurity AI agents better by guiding them strategically with context-additions that resemble how "humans think" using Game Theory? Our algorithm, Generative Cut-the-Rope (G-CTR) implements this and reduces ambiguity, collapses the LLM’s search space, suppresses hallucinations, and keeps the model tightly anchored to the most strategically relevant parts of the problem.
Cybersecurity AI: The World’s Top AI Agent for Security Capture-the-Flag (CTF)
Are Capture-the-Flag competitions obsolete? In 2025, Cybersecurity AI (CAI) systematically conquered some of the world's most prestigious hacking competitions, achieving Rank #1 at multiple events and consistently outperforming thousands of human teams. Across five major circuits have become a solved game for well-engineered AI agents.
Cybersecurity AI in OT: Insights from an AI Top-10 Ranker in the Dragos OT CTF 2025
Practical insights learned from participating in the Dragos OT CTF 2025 using the Cybersecurity AI (CAI) framework, highlighting strengths and limitations of AI agents when operating in real OT challenge environments and detailing actionable lessons for defensive and offensive workflows.
Cybersecurity AI Benchmark (CAIBench)
Existing benchmarks assess isolated skills rather than integrated performance. To address this limitation, we present the Cybersecurity AI Benchmark (CAIBench), a modular meta-benchmark framework that allows evaluating LLM models and agents across offensive and defensive cybersecurity domains, taking a step towards meaningfully measuring their labor-relevance.
Cybersecurity AI: Evaluating Agentic Cybersecurity in Attack/Defense CTFs
Empirical evaluation of AI systems in cybersecurity Attack/Defense CTFs reveals defensive agents achieve 54.3% patching success versus 28.3% offensive initial access, though operational constraints eliminate this advantage, providing first controlled evidence challenging AI attacker superiority claims.
Cybersecurity in Robotics: Challenges, Quantitative Modeling, and Practice
In cooperation with other researchers, this book stipulates the inclusion of security in robotics from the earliest design phases onward. We advocate for quantitative methods of security management, cover vulnerability scoring systems and account for the highly distributed nature of robots.
Securing robots in OT enviroments
We show how simple attacks are feasible in OT and how an industrial cybersecurity solution is not capable of capturing the complexity of modern robot interactions. We extend one of such solutions with a robot-specific Endpoint Protection Platform (EPP) and successfully protect the robot from attacks.
.png)