A recent comprehensive study spearheaded by researchers at the University of Cambridge shines a penetrating light on the rapidly evolving landscape of AI agents, unveiling a stark transparency deficit in safety documentation amidst their surging integration into daily life. Published in 2026, this landmark investigation—termed the 2025 AI Agent Index—evaluates thirty cutting-edge AI agents from global tech hubs predominantly in the United States and China, revealing profound gaps in safety practice disclosures despite the increasing autonomy and real-world impact of these systems.
These AI agents encompass a diverse range of functionalities, including conversational chatbots, autonomous web browsers, and enterprise automation tools designed to enhance productivity in various domains such as travel booking, online shopping, and corporate workflow management. While their proliferation promises unprecedented efficiency and assistance, the study unveils an alarming lag in safety transparency that could eventually undermine user trust and expose society to unforeseen risks.
The research team, incorporating prominent scholars from institutions like MIT, Stanford, and Hebrew University of Jerusalem, meticulously analyzed public data coupled with developer interactions. They identified that a mere four agents out of the thirty reviewed provide formalized safety documentation known as “system cards.” These cards serve as comprehensive dossiers outlining an AI agent’s autonomy, behavioral protocols, and importantly, detailed risk analyses. This paucity of documented safety evaluations indicates a troubling opacity that hinders thorough external assessment of potential vulnerabilities.
Most AI developers prioritize broadcasting their agents’ capabilities and performance features, yet significantly underreport safety-related information. The study quantifies this phenomenon as a “transparency asymmetry,” where information on operational prowess far outweighs disclosures of governance, benchmarking, or risk mitigation strategies. Twenty-five agents in the Index failed to reveal any internal safety assessment results, and twenty-three provided no evidence from independent third-party evaluations, critical components for establishing empirical trustworthiness in AI systems.
Particularly concerning are AI-enhanced web browser agents that interact autonomously with the open internet. Their design often includes mimicry of human browsing patterns and the ability to execute complex sequences such as clicking links, filling forms, and completing purchases on behalf of users. This class of agents demonstrated the highest autonomy levels alongside the greatest rates of missing safety-related information, with 64% of safety parameters unreported. The absence of established behavioral standards or robust disclosure mechanisms for these agents raises significant concerns over their unchecked influence on online ecosystems.
Moreover, the study highlights that several browser agents employ IP addresses and code structures specifically engineered to bypass anti-bot detection techniques, blurring the lines between human users and automated systems. Such indistinguishability challenges website operators’ capacity to regulate traffic and content integrity, potentially destabilizing digital marketplaces and content platforms dependent on accurate identification of legitimate users versus automated scraping or exploitation.
Chinese AI agents, though less represented in the Index, displayed a similar trend of sparse safety transparency; only one out of five examined disclosed any formal safety frameworks or compliance protocols. This lack of openness extends to critical aspects like prompt injection vulnerabilities, a mode of attack where manipulative inputs can override AI safeguards. This ability to covertly influence agent behavior underscores the urgency for rigorous safety assessments and public accountability.
Interestingly, foundational AI models like GPT, Claude, and Gemini underpin nearly all of the agents outside China, resulting in a systemic concentration of reliance on a few core architectures. While this affords efficiency in development, it simultaneously introduces potential single points of failure. An issue in any of these foundational models—be it safety regression, service interruption, or pricing adjustments—could cascade broadly, impacting hundreds of dependent AI agents, amplifying the scale of risks and necessitating coordinated safety oversight.
The study draws attention to the critical oversight by many developers who focus predominantly on foundational language model safety, neglecting the complex emergent behaviors that arise from agent-specific components such as planning modules, memory management, policy frameworks, and real-world interaction capabilities. Since these components critically shape autonomous agent behaviors, the dearth of disclosure in their safety evaluations represents a glaring gap in the current AI safety ecosystem.
One illustrative case is Perplexity Comet, an autonomous browser-based AI agent characterized both by high operational independence and pronounced opacity regarding safety practices. Marketed as functioning “just like a human assistant,” Comet has already garnered legal scrutiny for its failure to disclose its AI nature during interactions with services like Amazon, emphasizing the palpable risks of stealth AI operations in commercial domains without transparent safeguards.
Security researchers have previously exposed vulnerabilities where malicious web elements can hijack browser agents to execute unauthorized commands or exfiltrate private user data, revealing the fragile trust boundary between AI agents and their digital environments. The current study accentuates that without systematic safety evaluations and public scrutiny, such vulnerabilities may remain latent until exploited with potentially severe consequences in real-world contexts.
In conclusion, the 2025 AI Agent Index elucidates a crucial disconnect between the accelerating deployment and sophistication of agentic AI systems and the equally vital development of safety governance and transparency frameworks. As AI agents continue to gain autonomy and embed themselves deeper into everyday activities, the study urgently calls for standardized safety disclosure norms, comprehensive evaluations, and multi-stakeholder collaboration to mitigate systemic risks and harness AI’s full societal benefits responsibly.
Subject of Research: Investigation and documentation of technical capabilities and safety attributes of deployed autonomous AI agents.
Article Title: The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems
News Publication Date: 19-Feb-2026
Keywords: AI agents, AI safety, transparency, autonomous systems, system cards, AI governance, AI browser agents, prompt injection vulnerabilities, AI agent autonomy, foundational models, AI risk assessment, AI regulation
