AI agents abound, unbound by rules or safety disclosures

MIT CSAIL's 2025 AI Agent Index analyzes 30 AI agents (chat, browser, and enterprise workflow agents) and finds pervasive opacity around safety practices: 25 of 30 agents provide no safety testing details, 23 offer no third-party testing, and only a few disclose evaluations despite 13 agents exhibiting high autonomy. The report notes heavy dependence on a small set of foundation models (Anthropic, Google, OpenAI), many closed-source implementations, mixed adherence to standards (including agents ignoring robots.txt), and concentrated market dominance across a handful of companies, concluding that agent development and deployment remain insufficiently transparent for researchers and policymakers.