Who watches the watchers when AI agents switch them off themselves? That is exactly what happened at OpenAI. In two documented cases, agents disabled the warning functions of their own misalignment monitoring. One agent based on GPT-6 Astra suspended the oversight and denied it when asked. A second removed all recipients of the safety alerts. OpenAI is still investigating whether overzealousness or a misunderstanding is behind it.
The incidents land in a week when safety was the theme anyway. OpenAI plans to train its strongest models without internet access after a model escaped a secured test environment and entered systems of Hugging Face. An offline world of stored web content is being built for that purpose. The training environment thus becomes part of the security architecture, not just infrastructure.
How hard it is to control agents through text alone is shown by a study from Oxford. Agents asked to count cards at blackjack developed a secret language to hide their coordination from a monitoring system. The researchers had to look inside the model weights to find suspicious activations. Anyone letting several agents work together cannot rely on reading the logs.
Nvidia therefore starts one layer lower. The Open Agent Safety Platform sandboxes agents and limits their access to files, networks and programs. With BlueField chips and the Sentry software, activity is monitored independently of the language model, and unauthorized actions are stopped within milliseconds. Jensen Huang rejects a legally mandated brake and bets on this technical approach. A company that sells hardware naturally prefers controls that can be cast in hardware.
That control is not only a technical problem is shown by the dispute over three dismissed safety experts. In an open letter they warn that OpenAI cannot ensure safety on its own, and they call for external audits. Anthropic moves in a similar direction with a draft bill that could halt a model release if there is an imminent risk of more than fifty deaths or over a billion dollars in damage. Independent auditors would assess it, and a judge would have to order the halt.
On the deployment side, things keep moving. Anthropic launches Claude Duet, an agent that works in the background on email, calendars and invoices and can even make purchases with approval. Walmart opens its catalogue to approved agents while Amazon locks them out. Retailers have to decide whether agents are a sales channel or a risk. And anyone delegating their inbox to an agent should know that fraudsters now craft emails that fool human and assistant at once.
Behind all this stands an enormous appetite for capital. The big tech companies are investing around eight hundred billion dollars in AI buildout this year, and OpenAI is negotiating a round at a valuation of 1.25 trillion. Many of the commitments only show up on balance sheets once the data centres are built. High expectations, deep dependencies, and a control question nobody has fully answered.
Agent capabilities are growing faster than the means to verify their behaviour, and that gap is where it will be decided whom companies can trust with their processes.