AI agents abound, unbound by rules or safety disclosures

Summary

MIT CSAIL’s 2025 AI Agent Index inspects 30 agentic systems and highlights a striking lack of transparency around safety, evaluation and governance. The report classifies agents across six dimensions — legal, technical capabilities, autonomy & control, ecosystem interaction, evaluation, and safety — and provides per-agent annotation fields on its website. Researchers found most agents disclose little or no safety testing, many are closed-source, and most wrap a small number of foundation models from a handful of companies. The index warns these layered dependencies and sparse disclosures make real-world analysis and accountability difficult.

Key Points

MIT CSAIL analysed 30 AI agents in the 2025 AI Agent Index, down from 67 in 2024 but with deeper per-agent analysis.
Agents take multiple forms: chat apps with tools, browser-based agents, and enterprise workflow agents.
Safety disclosures are scarce: of 13 agents classed as having frontier autonomy, only four publish any agentic safety evaluations.
25 of 30 agents provide no public details about safety testing; 23 offer no third‑party testing data.
Most agents are closed-source (23/30); only seven open-sourced their agent frameworks or harnesses.
Agents commonly rely on a few foundation models (Anthropic, Google, OpenAI), creating hard-to-evaluate chains of responsibility.
The Index flags agents ignoring established web norms (robots.txt), showing traditional controls may be inadequate.
Market concentration is notable: a handful of companies dominate development, complicating governance and oversight.

Context and Relevance

The report matters because organisations, regulators and security teams need visibility into how agentic systems are built, tested and governed. With agents being deployed for tasks from email triage to potentially harmful uses, the lack of standardised safety reporting and the reliance on a small set of foundation models raise supply‑chain, accountability and security concerns. This intersects with ongoing debates about web scraping, automated code submission risks, and whether current internet protocols can curb unrestrained agent behaviour.

Why should I read this?

Short version: if you work with AI, run web properties, or worry about cyber risk, this is useful intel. MIT dug through 30 live agents and found that most makers talk up features but stay quiet on safety. It’s a tidy distillation of where the transparency gaps and concentration risks are — saves you poking through dozens of vendor pages yourself.

Source

Source: https://go.theregister.com/feed/www.theregister.com/2026/02/20/ai_agents_abound_unbound_by/