OUR RESEARCH
TAICI develops benchmarks and tools to measure and defend against AI-enabled manipulation, social engineering, and influence operations. Our initiatives quantify how AI systems can be exploited, and how they can exploit humans.
Measuring AI Uplift for Phishing and Scams
A benchmark measuring how much uplift AI models provide to phishers and scammers. ScamBench quantifies how effectively frontier models can be misused to plan, write, and run scam operations — standardized tasks with graded outcomes, comparable across models and versions.
Learn MoreMeasuring AI's capability, propensity, and vulnerability to manipulation
ManipulationBench is a behavioral benchmark for AI agents, measuring how susceptible frontier AI models are to social engineering and how capable they are of manipulating humans. ManipulationBench goes beyond binary pass/fail tests by measuring manipulation depth on a continuous scale across psychologically grounded, multi-turn attack scenarios, including agent-to-agent, human-to-agent, and agent-to-human tracks.
Learn MoreHow AI-enabled scams unfold, stage by stage
An interactive map of how modern scams unfold across nine stages — from defining the goal to laundering the proceeds — synthesizing five leading fraud and social-engineering frameworks, with an analysis of where current AI delivers the largest uplift to attackers.
Learn MorePractical Internet Safety for Everyone
Fred's Internet Guide is a public-facing resource for navigating the modern internet safely, covering scam awareness, phishing defense, voice-clone fraud, and OSINT-driven personal security. It pairs hands-on tools with plain-language guidance for everyday users.
Learn MoreTAICI works at the intersection of computer security, national security, and AI. Our mission is to reduce catastrophic risk from AI-enabled cyberterrorism, AI-driven totalitarian lock-in, and loss of control from advanced AI systems.
Our evaluation frameworks quantify AI security risks, such as how systems and users can be socially engineered or manipulated. Our work includes large-scale empirical studies (N ≈ 4,100) showing that AI already performs on par with human experts in phishing and voice scams, while dramatically reducing cost and increasing the scale of such attacks.
We collaborate with government bodies, leading tech companies, and frontier AI labs to translate our findings into policy actions and real-world defense mitigations.
Independent adversarial assessments of frontier AI systems, from jailbreak testing to multi-turn social engineering resistance. ScamBench and ManipulationBench provide standardized, reproducible evaluation at scale.
Quantifying how AI models can be socially engineered and how they can manipulate humans. Our ManipulationBench tracks measure manipulation depth across agent-to-agent, human-to-agent, and agent-to-human scenarios.
Preparedness exercises and rapid response for model misuse, prompt-injection and tool-use failures, data exfiltration, and emergent harmful behaviors.
Hardening the full stack around AI systems: data governance, access control, logging, provenance, and secure tool integration in high-trust environments.
Measuring and reducing human susceptibility to AI-driven deception, persuasion, and influence operations, informed by ScamBench social engineering data and ManipulationBench agent-to-human persuasion research.
Contributing benchmarks, scoring frameworks, and practical standards for safe AI, including the Manipulation Depth Index (MDI) and persuasion taxonomy used across our evaluation tools.