An air-gapped computer has no network cable and no Wi-Fi. Yet a 2024 paper showed that electromagnetic emissions produced by nearby RAM while it operates could encode secrets in memory as radio signals and carry them beyond the isolation boundary. The researchers did not break RAM; they exposed a blind spot in the assumption that physical isolation automatically guarantees security.
Original paper: arXiv:2409.02292
The paper, titled “Leaking Secrets from Air-Gap Computers by Spelling Radio Signals from RAM,” was published on arXiv in 2024 under the identifier 2409.02292.
The work also reached the Hacker News front page, earning 84 points and generating 40 comments. That discussion was lively, but it should not be interpreted as proof that “every air-gapped machine is compromised.” Attack distance, collection equipment, hardware models, and signal-processing methods all affect the results.
The Attack Makes Memory Speak Instead of Connecting to a Network
RAM emits electromagnetic radiation as it reads and writes data. The more pronounced the changes in current, the easier the resulting emissions may be to observe. Processors, memory controllers, and RAM continually switch states, producing weak but constantly changing electromagnetic signals nearby.
These signals are generally neither intelligible speech nor standard Wi-Fi broadcasts. Attackers re-encode them by mapping secrets to characters represented by radio waveforms, then recovering those characters from the collected emissions.
The paper calls this technique “spelling radio signals from RAM.” Its importance lies not in a new wireless protocol, but in transferring memory states from an isolated host across a physical boundary.
The complete path can be divided into four steps:
- The isolated computer runs a program, causing the target data to appear in RAM.
- The attacker uses a radio receiver to collect nearby electromagnetic emissions.
- The attacker analyzes the signal and identifies features produced by RAM activity.
- The attacker decodes the radio signal and reconstructs the secret.
This differs from conventional malware-based exfiltration, in which malware selects the data to transmit. The paper shows that even without a network interface, the hardware's ordinary operation can become an outbound channel.
Risks That Air-Gap Isolation Does Not Eliminate
Air-gapped deployments are commonly used in industrial control systems, critical infrastructure, intelligence processing, and high-security laboratories. The design rationale is straightforward: remove network paths and reduce the remote attack surface.
The problem is that data security depends on more than logical channels. Physical channels still exist around an isolated host:
- RAM and other circuitry emit electromagnetic signals into the surrounding space.
- Power lines may carry electrical noise related to the device's state.
- Keyboards, displays, cooling fans, and USB devices may reveal patterns of operation.
- Acoustic, optical, and thermal side channels may expose screen content, keystrokes, or computational state.
- Maintenance personnel may introduce risks through removable media, patches, or manual procedures.
In other words, an air gap blocks network communication, but it does not automatically remove the electromagnetic environment in which the device operates.
This does not mean operators should abandon air-gapped designs. They remain an important defense layer. The assumption that needs revision is:
“No network connection” means that data cannot leave through conventional network interfaces; it does not prove that the data cannot leak through other media.
If an attacker can approach the target, deploy a receiver, and process the collected signal, network isolation is no longer complete isolation.
Experimental Details From the Paper
The research targeted an air-gapped device and aimed to demonstrate that RAM-generated emissions could carry secrets across the isolation boundary.
According to the paper's abstract, the attack chain uses RAM electromagnetic emissions as its signal source, encodes data as radio signals, and recovers the content through collection and decoding. It is a hardware-side-channel and cross-domain data-exfiltration security case, not the release of a new model, framework, or AI tool.
Two easily confused concepts need to be separated.
First, the paper does not read plaintext directly from RAM. The attacker still needs the secret to appear in the target memory, or needs the program to generate electromagnetic features that can be recovered while it runs.
Second, this is not universal wireless eavesdropping. The experiment described in the paper depends on specific hardware, collection equipment, signal-processing methods, and environmental conditions. Its results cannot simply be extrapolated to reading arbitrary keys through several walls or from hundreds of meters away.
The work nevertheless exposes a blind spot common in deployed environments. Many security assessments check only whether a system has:
- An Ethernet port;
- A wireless network adapter;
- USB controls;
- Malware defenses.
Hardware side channels remind reviewers that the device remains active and the secret remains in memory. If someone nearby can collect the signal, the isolation boundary must be reassessed.
In high-value environments, the assessment question should shift from “Can this system connect to a network?” to “What physical-layer resources can an attacker access?” For example:
- Can radio receivers be deployed near the target?
- Do the RAM, CPU, and motherboard generate sufficiently distinctive emissions?
- Does the secret remain in memory for an extended period?
- Do power supplies, display outputs, or cooling systems create additional side channels?
- Does the isolation zone have radio-frequency detection and equipment controls?
The Boundaries of Confidential Computing Need Redrawing
The research has no direct connection to AI Agents, and it does not reveal a new vulnerability in any Agent framework. It is relevant, however, to the environments in which Agents operate.
Suppose an AI Agent is processing contracts, code, keys, or internal documents on an air-gapped server. To reduce risk, the team may enable a TEE, use memory encryption, and block access to external networks.
This combination is stronger than a bare-metal system, but several boundaries remain:
- A TEE generally protects data during computation. When an application reads data into RAM or writes it there, the threat from physical side channels does not automatically disappear.
- Memory encryption protects data at rest and in transit over the bus, but it does not prevent observable electromagnetic or electrical activity while a device is running.
- Agent tool use expands the set of components that handle data. Browsers, terminals, MCP services, and external devices may all become part of the data flow.
- If maintenance equipment can enter an isolation zone, network restrictions cannot prevent every physical-layer attack.
Confidential computing therefore should not be treated as a complete security conclusion. A better approach uses layered controls:
- Limit how long secrets remain in memory and where they are stored.
- Clear sensitive buffers once they are no longer needed.
- Isolate high-risk components that produce substantial radio-frequency or electrical noise.
- Apply equipment-access controls and radio-frequency activity monitoring within isolation zones.
- Determine whether secrets need to enter RAM at all or can exist only briefly during computation.
- Threat-model trusted execution environments, memory encryption, and endpoint controls separately.
The central point is not the headline-style claim that “RAM broadcasts secrets.” It is that security boundaries extend beyond software interfaces: they also cross silicon, power supplies, physical space, and operational procedures.
Editorial Perspective
The research changes the threat model rather than directly breaking every air-gapped system. Traditional security arguments often treated network isolation as the final boundary: without a network adapter or cable, there is no path for data to escape. The paper adds the hardware's physical layer. Even without actively connecting to a network, RAM produces electromagnetic emissions that form analyzable signals.
My expectation is that future assessments of high-security computing environments will more often ask whether radio-frequency collection equipment can be deployed nearby and how long secrets remain in memory. TEE and memory encryption remain worthwhile, but neither can carry an entire isolation strategy on its own.
This differs from previous malware-based exfiltration methods. The attacker does not need to control software on the target or compromise a network protocol. The attack exploits hardware activity that necessarily accompanies computation. Its applicability across different RAM modules, motherboards, and receiver distances remains uncertain, so the findings should not be interpreted as evidence that air-gapped deployments have failed. The more accurate conclusion is that physical isolation remains useful, but it is not absolute offline security; it requires supporting physical-layer monitoring and strict management of isolation zones.