Opinion: OpenAI says advanced AI model could talk across air‑gapped computers via thermal signals

In what is a truly nightmarish, Terminator-type scenario, OpenAI now thinks a determined, sufficiently advanced AI model can communicate across two completely isolated - albeit proximal - computers by using temperature changes as a morse code of sorts.

OpenAI's Noam Brown: "We never want to be in a situation again where we underestimate the AI"

In what appears to be a scene ripped straight out of a dystopian sci-fi movie, OpenAI's Noam Brown has just demolished an oft-touted measure against runaway AI capabilities: air-gapping computers.

For the uninitiated, air-gapping means physically isolating a computer - or a network of computers - so it has no network connections or data links to any other system, thereby creating a hard barrier against AI-driven exfiltration.

In what is truly scary, however, Brown references studies to illustrate just how far a determined AI model can go to escape its physical prison, so to say:

"There are studies, and this is mostly academic, where you can have two computers next to each other that are air-gapped and they're still able to communicate with each other because they have temperature sensors."

Basically, modern CPUs, GPUs, and motherboards have built-in thermal sensors such as temperature diodes. These exist primarily for thermal management, allowing the system to spin up fans, throttle clocks, or shut down if it overheats. Some studies, however, have demonstrated that it is possible to maintain a very low-bandwidth covert channel between two air-gapped computers, who can talk to each other by running one of the CPUs hot and then using the ensuing thermal variations as a code that the other computer can interpret.

For OpenAI, such tail-risk vectors still pose a sizable headache, especially as the AI lab feels that chain-of-thought monitorability is already degrading as models become smarter at concealing their true inner workings.

Of course, all of this brouhaha around AI alignment and safety truly kicked off earlier this summer when OpenAI's cybersecurity agents broke out of their isolated test environments by exploiting an undiscovered vulnerability, and then went on to attack the model repository Hugging Face in their quest to find a solution to a cybersecurity-related eval.

Of course, some have pointed out that in that incident, AI agents could obtain software through an internet-enabled intermediary, which suggests that they were not completely isolated. What's more, those agents were primed not to give up, which explains the doggedness with which they pursued a potential solution.

The conversation around AI alignment, however, truly picked up after Anthropic's Dario Amodei published an open letter last weekend, committing the high-flying AI lab to third-party evaluations, most likely via embedded personnel from Model Evaluation & Threat Research (METR). Amodei also called for a coordination on a global level to pace the progress on AI, with OpenAI, Elon Musk, and Microsoft all communicating their acquiescence in short order.

Also, just a few hours back, some researchers were able to use Anthropic's Claude Opus 5 to "compromise multiple OpenAI employees’ ChatGPT accounts," to gain access to "internal OpenAI repositories, and potentially many other connectors."

While almost everyone agrees that AI alignment is shaping up to be a critical issue, skepticism persists, with only 20 percent of the enterprise customers in a recent survey interpreting the proposed slowdown of AI development as "genuine safety measures."

Of course, as we explained in a dedicated post recently, much of this skepticism stems from the self-serving nature of some of the steps proposed by Anthropic and OpenAI. For instance, Anthropic's preferred choice to run third-party evaluations, METR, is a non-profit with hefty links to Anthropic itself, replete with a revolving door of sorts between the two organizations when it comes to safety researchers.

What's more, embedding third-party non-profits within alignment-related workflows might also be a clever workaround for collecting and training on high quality expert traces, funded by non-profit dollars, with the resulting models made available without public deployment and the attendant margin-destroying distillation threat.

Moreover, when Anthropic calls for an explicit role of the government in dictating how the AI industry evolves, that's regulatory capture, one that can and will reinforce the Anthropic-OpenAI duopoly.

Finally, for Anthropic and OpenAI, pacing can work wonders to boost their margins by extending the useful life of their reigning workhorse models, thereby reducing the attendant R&D amortization costs.

Despite these skepticism-inducing factors, the scenario that OpenAI's Brown has just painted is quite scary, one that merits a thorough and informed discourse on a viable approach to AI alignment.

Follow Wccftech on Google to get more of our news coverage in your feeds.