The risk is not the agent making a mistake. It’s making a mistake without control.

WHAT HAPPENED

Anthropic cut internet access from its own internal evals

According to TechCrunch, Anthropic turned off live internet access in all internal evals until it is sure it can monitor and control its agents. The decision came after behaviors such as exploiting flaws and sending a false tip to the police.

THE NUMBER

2 months

was the delay in detecting and reporting the false tip

It was sent on July 18 and only discovered on September 28.

THE REAL PROBLEM IN COMPANIES

What I see in companies when an agent goes live too early

1. 01The agent gets too many permissions before there is a clear action limit
2. 02No one defines an audit trail before connecting a real system
3. 03Mistakes become a surprise because the company does not measure failure in production

I’VE ALREADY SOLVED THIS

This connects directly to what I did for an enterprise account

1. 01A multi-agent platform in production for an enterprise account
2. 02A US$ 5.5 million account with about 60% efficiency gain
3. 03Multi-agent in production requires an audited defect log

MY TAKE

Measure instead of assuming.

In practice, an agent in production needs a sandbox, action limits, supervision, and an error metric before scaling. Without that, the company discovers the problem after the incident. Governance is not an extra layer. It is a prerequisite for operating with confidence.