OpenAI Researcher Shows Heat Emissions Can Exfiltrate Data From Air‑Gapped PCs

A researcher at OpenAI has shown that even computers that are intentionally isolated from any network—known as air‑gapped systems—are capable of leaking data by varying their own heat emissions, confirming prior academic theories about thermal covert channels.

Air‑gapped devices form the backbone of highly secure settings, ranging from governmental laboratories to essential infrastructure, because they lack any direct electronic connection to outside networks. The new test demonstrates that by intentionally altering processor load, a machine can produce minute temperature changes that a nearby sensor can detect. These variations can be modulated to encode bits, though the transmission speed remains modest.

In the demonstration, the OpenAI‑provided hardware generated a patterned heat signal while a second unit, positioned a few feet away, captured temperature shifts using an off‑the‑shelf thermal sensor. The resulting transmission was enough to convey short strings such as cryptographic keys or passwords, proving that the channel, while slow, is technically feasible.

This finding expands the growing list of side‑channel attacks that leverage unconventional emissions—acoustic, electromagnetic, optical—to bridge air gaps. Security professionals point out that the risk is not immediate for most entities, as exploiting the method demands close physical proximity and precise timing. Nevertheless, sites handling highly classified data may need to revisit physical‑security rules, for example by enforcing minimum distances between isolated hardware and any temperature‑monitoring devices.

Existing countermeasures include watching for abnormal CPU load patterns, applying thermal shielding, and limiting the placement of temperature sensors near critical equipment. Analysts say the broader takeaway is that defense must stay layered, covering both traditional network safeguards and atypical physical side channels. Although the heat‑based method is unlikely to spark widespread alarm, it highlights the need to keep threat models current as researchers reveal new ways isolated systems can inadvertently “talk.”