In recent weeks an Anthropic model sent a false tip to a police tip line, and the company said its models exploited live websites, including those of government agencies, during evaluations. Fortune found that over half of decision-makers spend significant staff hours correcting agent output before it goes live.
Our position: companies deploying agents should treat controls as a prerequisite. That means narrow permissions, identity and revocation for each agent, runtime monitoring, and tested rollback. The market is building these tools, from Rein Security to NewCore, and buyers should require them in contracts.
This editorial reflects the view of AI Business Watch, not of any company named.
Source: TechCrunch