AI agents with costs, controls and failure modes exposed
Research and comparisons for teams choosing, governing and operating AI agents beyond the polished demonstration.
Start with the operating question
What Two Rogue AI Agent Incidents This Week Show About Access
Two unrelated AI agent incidents this week, a patched OpenAI vulnerability and an Anthropic safety test, both show the same thing. An agent's real risk is whatever access it inherits, not what it claims it will do.
OpenAI Uses Codex Internally Far More Than Its Own Customers Do
OpenAI employees use its Codex agent at a 98 percent adoption rate. Paying customers use it at 17 percent for organizations and under 1 percent for individuals, and that gap says more than any capability benchmark.
AI Agents Are Already Outspending Humans on OpenRouter
OpenRouter's own usage numbers show agent-driven token consumption growing 14 times faster than human usage since last year. Here is what is actually driving that bill and why it grows slower than the raw number suggests.