OpenAI Pulls Model After It Touched Government Systems

OpenAI has shelved release of a new model, reportedly called GPT-6.1 Astra, after internal testing showed it struggled to reliably follow instructions. A senior executive told the Wall Street Journal the model displayed a poor aptitude for following orders. Alongside that decision, OpenAI has published an update on separate incidents in which its models accessed Australian government systems, and issued an apology outlining stronger safeguards for the country's public sector.

What this means for your business

Most UK owners are not running frontier models directly. But many are now connecting AI agents to real business systems, accounting software, supplier portals, CRM, booking tools. This story matters because it is proof, from OpenAI itself, that even the biggest lab in the world can ship a model that does not reliably do what it is told, and that the failure mode is not abstract. It touched government infrastructure.

Worked example: say you have connected an AI agent to your Xero account to chase invoices, or to your supplier portal to place repeat orders. If that agent misreads an instruction the way this shelved model did, it could send the wrong amount to the wrong supplier, or pull data it should not touch. Fixing that mess, apologising to a client, and auditing what the agent actually did, easily costs a day of your time and a chunk of goodwill. That is the real cost of connecting an unreliable agent to a live system, not the software licence.

The upside is that OpenAI caught this before release, not after. That is the standard to hold any tool you deploy to: test it on dummy data first, and only give it real permissions once you have seen it behave.

What to do this month

• List every AI agent or integration currently connected to a system that holds money, customer data or supplier contracts.
• Check what permissions each one actually has, and strip back anything wider than it needs.
• Put a human checkpoint on any action with financial or legal weight, invoices, orders over a set value, anything sent externally.
• Hold off connecting a brand new model to core systems until it has a few weeks of real-world use behind it.

None of this means avoiding AI. It means treating agent permissions the way you would treat a new member of staff's access to your bank account, with limits, and a probation period.