What would it take for someone to make your AI model lie to you — silently, instantly, and without leaving a trace in the software? According to new research from Portland-based cybersecurity firm Eclypsium, the answer is uncomfortable: access to a server’s network, a few disabled security settings, and less than sixty seconds.
That finding, shared exclusively with CNN ahead of publication, reframes the entire AI safety conversation. The dominant anxiety about AI systems — whether they hallucinate, whether they pursue misaligned goals, whether their developers can be trusted — assumes the model itself is the attack surface. Eclypsium’s research points one layer deeper: to the chips, memory, and data center servers the model runs on.
The Reading
What Eclypsium Actually Did
Eclypsium’s team — staffed largely by engineers who previously worked at hardware firms including Intel — demonstrated the attack on an Nvidia computing platform. Their method exploited access to either a data processing unit (DPU), the system-on-chip that helps orchestrate data center operations, or a baseboard management controller (BMC), the small embedded computer used to manage servers remotely. BMCs are a long-standing concern in enterprise security: this summer, longtime cybersecurity researcher HD Moore disclosed multiple BMC vulnerabilities, and those same vulnerabilities serve as the entry point in Eclypsium’s AI attack chain.
Once on the network with the right access, an attacker can read and modify data in the machine’s main memory. That memory includes the “weights” — the numerical parameters that determine how an AI model interprets and responds to any given input. Changing those weights changes the model’s behavior. The modification takes less than a minute. The person querying the AI sees nothing unusual.
The consequences Eclypsium demonstrated were not abstract. In one scenario, the researchers manipulated a model to substantially increase the probability it assigned to a Chinese military invasion of Taiwan — a deeply consequential distortion in any national security or intelligence context. In another, an AI model running on tampered hardware told a user describing chest pain to stay at home rather than seek hospital care. The researchers say state actors — including China, Russia, and the United States — have the technical capability to carry out targeted versions of this attack.
Nvidia’s Response and the Limits of Existing Defenses
Nvidia, whose hardware was used in the demonstration, said in a statement that it had reviewed the hypothetical scenarios and “did not identify anything the industry wasn’t previously aware of.” The company noted that strong memory access protections can block the described attack path and that it welcomed working with Eclypsium on the research. Nvidia last month announced a new security platform designed to guard against hardware attacks and keep AI agents from going rogue, a platform Eclypsium acknowledges offers more security monitoring for AI agents.
But Eclypsium’s response to Nvidia’s defense is pointed: the effectiveness of those security controls “depends on how they are configured and used.” In many real-world deployments, those protections are not configured correctly — or are disabled entirely to preserve performance or compatibility. The attack scenario, the firm is careful to note, is not unique to Nvidia and could apply broadly across hardware vendors and the firmware that runs on them.
Two independent hardware security researchers who reviewed the Eclypsium findings at CNN’s request confirmed their credibility. Joe FitzPatrick, a hardware security instructor and researcher, said: “The attacks are fundamentally the same as any other direct memory access attacks and I don’t see any reason to doubt their findings or results.” Dominic Rizzo, senior fellow at the nonprofit Institute for Security and Technology, described the attack path as “pretty general” and argued that “the solutions have to start much deeper in the stack.”
Eclypsium executives said they briefed the National Security Agency and the Cybersecurity and Infrastructure Security Agency — the two agencies responsible for protecting military and civilian federal computer networks, respectively — on their findings.
The Broader Context: A Safety Debate Built on the Wrong Foundation
The timing of Eclypsium’s research is pointed. The race for AI supremacy between the United States and China has sharpened considerably in 2025, and real-world incidents are already illustrating how AI outputs can carry national security consequences. CNN reported earlier this year that a false AI-assisted US intelligence report — generated when a military analyst queried a chatbot about a ship’s manifest — nearly prompted the US military to board a vessel it incorrectly believed was carrying nuclear components. Whether that chatbot was a commercial product or a government system was never confirmed, but the episode illustrated that the gap between an erroneous AI output and a kinetic response is narrower than most people assumed.
Against that backdrop, President Trump and Chinese leader Xi Jinping discussed AI during a meeting last month, but the session produced no substantive accord on the technology. Trump has publicly resisted calls for guardrails on advanced AI, citing the risk that the US would fall behind China in the technology race — a framing that has historically prioritized model capability over infrastructure security.
What the Eclypsium research exposes is a structural blind spot in how the AI industry has organized its safety efforts. The dominant framework — model alignment, output filtering, red-teaming of model behavior — assumes that if the model itself is trustworthy, the system is trustworthy. But as the FTC’s broadening investigation into AI agents and incidents of AI agents behaving unexpectedly on government websites have made clear, the layers below the model — the hardware, firmware, and network stack — have received far less systematic scrutiny. Eclypsium’s finding suggests this is not merely an oversight but a compounding vulnerability: the more powerful AI models become and the more sensitive the decisions they inform, the higher the stakes of an undetected hardware-level compromise.
Edge Cases and Caveats
The attack is not trivially easy. It requires network access to a DPU or BMC, certain security configurations to be disabled, and a degree of technical knowledge — although Eclypsium notes, with some irony, that AI models themselves can help an attacker acquire that knowledge. In tightly locked-down, air-gapped environments with properly configured memory protections, the attack path described is significantly harder to execute. The question is how many real deployments of AI infrastructure — including those handling sensitive government or medical queries — actually meet that security bar.
Eclypsium’s CEO Yuriy Bulygin put the challenge plainly: “We can’t trust the model if we can’t trust the infrastructure it’s running on.” Senior vice president John Loucaides framed it as a question of incomplete attention rather than negligence: “Everybody’s talking about model safety, which is perfectly fine. Please do that. I just am trying to make sure that we don’t forget about all the computers it’s running on.”
How AI Hardware Attack Exposure Compares Across Deployment Models
Not all AI deployments carry equal exposure to the class of attack Eclypsium describes. The table below compares three common deployment configurations against the key risk factors — based on publicly known properties of each architecture, not vendor claims.
| Deployment Model | BMC/DPU Network Exposure | Memory Protection Configurability | Likely Attack Difficulty | Example Context |
|---|---|---|---|---|
| Public cloud (hyperscaler-managed) | Low — managed by cloud provider; tenant cannot access BMC | Provider-controlled; generally consistent | High — attacker would need to compromise hyperscaler layer | OpenAI API, AWS Bedrock |
| Enterprise on-premise / private data center | Medium-High — BMC often network-accessible by IT staff; configurations vary widely | Variable — frequently misconfigured or partially disabled for performance | Medium — requires insider or network access; config inconsistency is common | Self-hosted LLM, private inference cluster |
| Government / sovereign AI infrastructure | Potentially Low — air-gapped ideally, but supply-chain and insider risk remain | Should be strict; real-world adherence varies | Medium — state-actor adversary assumed; supply-chain attack vector relevant | Military intelligence tools, national AI platforms |
The implication is that enterprises running their own inference clusters — a group that is growing rapidly as AI infrastructure investment accelerates toward the trillions — carry the highest practical exposure, precisely because they control their own configuration and may not have dedicated hardware security expertise on staff.
Your Next Three Moves
Audit Your BMC and DPU Network Access Today
Map which systems in your AI inference stack have network-accessible baseboard management controllers or data processing units, and verify that access is restricted to a dedicated out-of-band management network with strong authentication. If BMC access is on the same network segment as production workloads, that is your most urgent remediation priority.
Validate Memory Protection Configuration on Every Inference Node
Confirm that memory access protections — including IOMMU settings and relevant firmware-level controls — are enabled and correctly configured on every server running AI inference workloads. Treat any node where these settings were disabled for performance tuning as a potential compromise vector until reviewed, and establish a configuration baseline you can monitor for drift.
Add Hardware Integrity to Your AI Risk Register
If your organisation has an AI risk framework, add hardware and firmware integrity as an explicit control domain alongside model alignment and output monitoring. Brief your security operations team on the BMC vulnerability class documented by HD Moore and referenced in the Eclypsium research, and track whether your hardware vendors are issuing firmware updates in response to disclosed vulnerabilities.











