OpenAI Researcher Shows Heat Emissions Can Leak Data From Air‑Gapped Computers

A scientist at OpenAI has proved that even computers intentionally isolated from any network—known as air‑gapped systems—are capable of exfiltrating data by altering their heat emissions, confirming prior academic theories on thermal covert channels.

Such air‑gapped devices form the backbone of high‑security settings, ranging from government research facilities to essential infrastructure, precisely because they have no direct electronic connection to outside networks. The recent test demonstrates that by intentionally adjusting the CPU’s workload, a machine can generate minute temperature changes that a close‑by instrument with a basic thermal sensor can sense. These variations can be modulated to send binary information, though only at a modest speed.

For the demonstration, the scientist employed ordinary OpenAI‑provided hardware to emit a structured heat pattern, while a second unit positioned a few feet away captured the temperature shifts using a commercial off‑the‑shelf sensor. The captured stream was capable of transmitting brief strings—like cryptographic keys or passwords—showing that, despite its low bandwidth, the channel is practically feasible.

This finding expands the expanding list of side‑channel exploits that leverage unconventional emissions—acoustic, electromagnetic, or optical—to cross air gaps. Analysts point out that the danger is not imminent for most entities, since a successful attack would demand close physical proximity and precise timing. Still, sites dealing with highly classified data might have to revisit physical security guidelines, for example by enforcing minimum spacing between isolated machines and any temperature‑sensing equipment.

Countermeasures are already available: watching for unusual CPU load signatures, applying thermal shielding, and limiting the proximity of temperature sensors to vital hardware. Analysts say the overarching takeaway is that security must stay multi‑layered, addressing both standard network safeguards and atypical physical side channels. Although a heat‑based conduit is unlikely to spark widespread alarm, it highlights the need to constantly revise threat models as researchers reveal fresh methods for isolated systems to unintentionally “communicate.”