AI Agents Aren’t Waiting for Global Safety Rules

What it is

A Galaxy Research newsletter analysis built around three data points on agentic AI deployment that the UN safety proposals from Sam Altman (OpenAI) and Dario Amodei (Anthropic) don't address: the Australia Medicare portal incident, Anthropic's autonomous DNA-database search, and token-usage economics showing agents overtaking humans.

How it works

The argument is economic, not technical: coordination proposed by frontier labs only binds the systems those labs control, while usage is shifting away from them. The page cites specific figures to support this:

• On OpenRouter, agent token consumption has exceeded human consumption since February 2026, growing ~14x from February to August versus 2.8x growth for human usage.
• Cloudflare (sitting in front of roughly a fifth of the web) originally forecast automated traffic would overtake human traffic by end of 2027, revised that to early 2027, then CEO Matthew Prince announced in June that the crossover had already happened.
• On Vercel's AI Gateway, average price per token fell 23.2% in August alone and is now less than half its level five months earlier.
• Open-weight models accounted for 56% of token volume on Vercel's gateway, up from 7% in December 2025.

Preconditions / Affected

The Vercel figures are explicitly scoped to traffic on Vercel's own gateway only, not the broader market — the article flags this itself. The other metrics (OpenRouter, Cloudflare) are platform-specific too, not industry-wide census data.

What to do

The piece doesn't prescribe actions for practitioners; its point is that frontier-lab self-regulation (slower releases, more evaluations, better incident reporting — the measures Altman and Amodei can actually commit to) addresses a shrinking share of total agent activity as cheap, open-weight models take over inference volume. The implied takeaway for security teams: governance and monitoring need to target where agents can act and what they can access, not which lab trained the model, since open-weight/non-frontier systems are an increasing majority of agentic token volume.

Caveats

The 56% open-weight figure and the 23.2% price drop are drawn from a single gateway (Vercel) and explicitly caveated by the article as not representative of the entire market. The Cloudflare crossover claim rests on a single executive statement (Prince, June) rather than published data in this piece.