6 hours ago
5
Systems
Intel ME and AMD PSP: The silicon layer nobody certifies
FEATURE Can digital sovereignty exist on American silicon?
Europe is pouring more than €2 billion into sovereign cloud initiatives designed to reduce exposure to US legal reach. The EU's IPCEI-CIS program funds infrastructure development. France qualifies operators under SecNumCloud, a framework with nearly 1,200 technical requirements promising "immunity from extraterritorial laws."
But most datacenters and qualified cloud operators still rely heavily on Intel or AMD processors. And inside those processors sits a computer beneath the computer: management engines operating at Ring -3, below the operating system, outside the control of host security software, persistent even when the machine appears powered off. Under the US Reforming Intelligence and Securing America Act (RISAA) 2024, hardware manufacturers count as "electronic communications service providers" subject to secret government orders.
Europe's frameworks certify the clouds. They don't assess the silicon.
The computer your OS can't see
That computer beneath the computer has a name. On Intel processors, it is the Management Engine (ME), or more precisely the Converged Security and Management Engine (CSME). On AMD, it is the Platform Security Processor (PSP). Both run at what security researchers call Ring -3, below the operating system, below the hypervisor, in a privilege level the host cannot see or log.
"It's a computer inside your computer," explains John Goodacre, Professor of Computer Architectures and former director of the UK's £200 million Digital Security by Design program. He is clear about what that means in practice. The ME has its own memory, its own clock, and its own network stack, and because it can share the host's MAC and IP addresses, any traffic it generates is indistinguishable from the host's own traffic to the firewall.
The architecture is not theoretical. Embedded in the Platform Controller Hub, the CSME is a separate microcontroller that operates independently of the host, with direct memory, device access, and network connectivity the host operating system cannot monitor. AMD's PSP works the same way.
Intel's Active Management Technology (AMT), the remote management feature the ME enables, exposes at least TCP ports 16992, 16993, 16994, and 16995 on provisioned devices. Goodacre notes that an attack surface exists on unprovisioned hardware too. These ports deliver keyboard-video-mouse redirection, storage redirection, Serial-over-LAN, and power control to administrators managing fleets of devices remotely. The capability has legitimate uses. It also provides a channel that operates at a level below what European sovereignty frameworks can attest.
Microsoft documented in 2017 that the PLATINUM nation state actor used Intel's Serial-over-LAN (SOL) as a covert exfiltration channel. SOL traffic transits the Management Engine and the NIC sideband path, delivered to the ME before the host TCP/IP stack runs. The host firewall and endpoint detection saw nothing, and any security tooling running on the compromised machine itself was equally blind. PLATINUM did not exploit a vulnerability. It exploited a feature, requiring only that AMT be enabled and credentials obtained. In documented cases, those credentials were the factory default: admin, with no password set.
Goodacre catalogues this and related scenarios in a 37-page risk assessment prepared for CISOs evaluating Intel vPro hardware connected to corporate networks. Its conclusion is blunt: connecting an untouched-ME device to corporate resources "exposes the organization to a class of compromise that defeats the host security stack in its entirety."
The ME does not stop when the machine appears to. Users recognize the symptom: a laptop powered off and stored for weeks is found, on next boot, to have a depleted battery. On modern thin and light platforms, what Microsoft documents as Modern Standby means "off" does not correspond to "all subsystems unpowered." The system-on-chip components the Management Engine runs on remain in low-power states, drawing enough to drain a 55 Wh battery over weeks, on the order of 100-200 mW continuous draw.
The implication is documented in Goodacre's risk assessment: "Whether the radio is in a Wake-on-Wireless-LAN listening state is firmware policy. On a device whose firmware has been tampered with during transit through the supply chain, the answer cannot be inferred from the visible power state." A laptop that appears off, in a bag, can associate with a hostile network the user has no knowledge of.
Professor Aurélien Francillon, a security researcher at French engineering school EURECOM, has spent years studying exactly this class of problem. Working with colleagues, he built a fully functional backdoor in hard disk drive firmware [PDF], a proof of concept demonstrating how storage devices could silently exfiltrate data through covert channels. Three months after presenting it at an academic conference, the Snowden disclosures revealed the NSA's ANT catalogue, which documented an identical capability already deployed in the field.
"The NSA were already doing it," Francillon says flatly. "Quite amazing." That background informs his assessment of the ME. "Yes, it can probably be used as a backdoor, like many other things, including BMC [baseboard management controller] and many other firmwares," he says. The question, he argues, is not whether the backdoor exists but whether operational controls make it unreachable in practice.
AMD faces the same architectural question. On April 14, 2026, researchers demonstrated the Fabricked attack against AMD's SEV-SNP confidential computing technology, achieving a 100 percent success rate with a software-only exploit. The Platform Security Processor proved vulnerable to the same class of compromise.
On server hardware, the picture is the same. Intel ME runs on servers under a different name, Server Platform Services or SPS, and the BMC, the remote administration controller standard in datacenter hardware, relies on it. "More or less the same," Francillon says of the server variant. For datacenter operators, he sharpens the focus further: "If I look at cloud systems, servers, I would be more concerned with the BMC," pointing to published research demonstrating remote exploitation of BMC vulnerabilities that could allow an attacker to reinstall or fully compromise a server. The BMC is not a separate concern from the ME: on server hardware, it is the primary network entry point into the SPS, making it both the most exposed interface and the most consequential.
Both Intel and AMD processors contain management engines that operate below the operating system. The silicon is designed by American companies and subject to American legal process.
The backdoor the CLOUD Act doesn't use
That legal process has teeth that most European policymakers underestimate. The CLOUD Act, passed in 2018, gave US authorities extraterritorial reach to data held by American companies. FISA Section 702 allows intelligence agencies to compel US persons and companies to provide access to communications. Both are well known in European sovereignty discussions. They operate through the front door: a legal order served on a company that controls data. Less well known is RISAA 2024, a law that opens a different entrance entirely.
RISAA amended FISA's definition of "electronic communications service provider" in ways that go beyond cloud operators and platform companies, and beyond the bilateral agreements that European policymakers have built their legal defenses around. Hardware manufacturers now fall within scope. Intel and AMD can be compelled, via secret orders with gag clauses, to cooperate with US intelligence access.
The mechanism through which that access could be exercised is the management engine: a persistent, privileged, network-connected runtime that operates below anything the host operating system can see or block. A SecNumCloud-certified operator can be legally isolated from American data demands. The processor inside its servers cannot. "You've actually got a policy mechanism by which any such machine anywhere can deliver any of its information," Goodacre says.
RISAA's two-year term expired on April 20, 2026, but Congress extended it by 45 days while debating reforms. Whether it is renewed, amended or allowed to lapse, the architecture it targets does not change.
SecNumCloud's blind spot
France's SecNumCloud is Europe's most rigorous attempt to build a cloud certification that is legally immune to American law. It did not emerge from nowhere. ANSSI, France's national cybersecurity agency, was established in 2009 as part of a broader effort to build institutional muscle on digital sovereignty long before the term became fashionable. When Edward Snowden revealed the scale of NSA surveillance in 2013, France's response was technical rather than rhetorical: ANSSI published the first SecNumCloud framework in July 2014. A decade later, that framework has grown to nearly 1,200 technical requirements.
At the time, SecNumCloud was a cybersecurity qualification, not a sovereignty instrument: it set requirements for architecture, encryption standards, access controls, and incident response, but said nothing about who controlled the underlying infrastructure or whose laws applied to it. The CLOUD Act changed that. Passed in 2018, it gave American authorities extraterritorial reach to data held by US companies, and suddenly a French cybersecurity framework had a geopolitical dimension it was not designed for. Version 3.2, introduced in 2022, added Chapter 19: a set of explicit requirements targeting extraterritorial law, mandating that only EU operators could run the service, that no non-EU party could access customer data, and that the provider could operate autonomously without external intervention. It promised "immunity from extraterritorial laws."
In December 2025, S3NS, a joint venture between French defense and technology group Thales and Google Cloud, operating Google Cloud Platform technology under French control, became the first "hybrid" cloud to receive SecNumCloud qualification. The certification triggered heated debate: was this real sovereignty, or American technology with a European flag?
But the debate missed a more fundamental question. Does SecNumCloud's certification reach as far as the silicon it runs on? Francillon is positioned to see both sides of that question. He sits on the French Technology Academy's working group on cloud security, a body that advises on the technical foundations of frameworks like SecNumCloud. And he has spent years studying firmware backdoors in academic literature and demonstrated them in practice.
He knows what the hardware can do, and he knows what the certification requires. His starting point is that SecNumCloud provides genuinely valuable protection, and that the silicon gap does not negate that. When asked whether SecNumCloud explicitly addresses Intel Management Engine or AMD Platform Security Processor vulnerabilities, his answer is unambiguous: "There is no direct requirement for firmware backdoor prevention."
The framework is not designed to be a technical specification for hardware-layer security. "The document aims to be generic and not dive into technical details," Francillon says. "Most of it is organizational security." What SecNumCloud does require is that providers build a proper threat model, consider mitigation mechanisms, and monitor administration gateways where external tech support could be exploited. The hardware layer was not addressed by oversight. It was left out by design.
Francillon's assessment is not a fringe view. Vincent Strubel, the director of ANSSI, the very agency that designed and administers SecNumCloud, is equally explicit about what the framework does and does not cover. In a January 2026 LinkedIn post addressing SecNumCloud's scope, he writes that all cloud offerings, hybrid or not, depend on electronic components whose design and updates are not 100 percent controlled in Europe. If Europe were ever cut off from American or Chinese technology, he argues, the result would be a global problem of security degradation, not just in hybrid clouds, but everywhere.
Strubel frames SecNumCloud carefully: it is "a cybersecurity tool, not an industrial policy tool." It protects against extraterritorial law enforcement and kill-switch scenarios. It was never designed to eliminate technology dependencies at the hardware layer, and no actor, state, or enterprise fully controls the entire cloud technology stack anyway.
One technology frequently cited in sovereignty discussions is OpenTitan, Google's open source secure element deployed on its server hardware and used within the S3NS infrastructure. Francillon is clear about what it is and, critically, what it is not. "OpenTitan is a secure element, a small chip on the side that can be used for protecting sensitive keys, providing signatures, making attestations," he explains. "It's a bit like a TPM." What it is not is a replacement for the main processor. "The Linux and all your applications will not run on it." OpenTitan sits alongside x86 infrastructure as an external root of trust, independent of the ME. That matters because the default embedded TPM lives inside the ME, making it subject to the ME attack surface. OpenTitan sits outside that boundary. The two address different problems entirely, and conflating them, as sovereignty advocates sometimes do, obscures where the residual exposure actually lies.
ANSSI's own technical position paper [PDF] on confidential computing, published in October 2025, concludes that Intel SGX, TDX, and AMD SEV-SNP are "not sufficient on their own to secure an entire system, or to meet the sovereignty requirements of SecNumCloud 3.2." Physical attackers are "explicitly out-of-scope" of vendor security targets. Supply chain attackers are "explicitly out-of-scope." The ME attack surface discussed in this article falls into neither category: it is a remote network threat, not a physical one. The paper's conclusion for users concerned about hostile cloud providers is stark: "Switch to a cloud provider they trust, or use their own hardware with physical security protection measures."
The castle with a structural flaw
Francillon does not dispute that SecNumCloud leaves the ME unassessed. His argument is that this does not matter in practice. "What I mean is that if there is a backdoor to access a room, it cannot be directly used if the room is in a castle. You have to pass the castle walls first." Network isolation, monitoring, and threat modeling are the walls. SecNumCloud's operational requirements mandate that administration gateways be isolated, that external tech support be monitored, that network segmentation prevents lateral movement. The Management Engine backdoor may exist, but the framework makes it unreachable except in what Francillon calls "very high-end attacks."
That qualifier matters. Francillon is not claiming perfect security. He is claiming that proper operational controls reduce the threat to a level where only nation state actors with significant resources could exploit it. For most threat models, he argues, that is sufficient. "Saying it is useless to do SecNumCloud because there is ME, or whatever backdoor in some hardware we don't control, is a mistake," he says. SecNumCloud improves security over deployments without such controls, he argues, provided that hardware is carefully evaluated and firmware securely configured.
The castle walls have a structural flaw that Goodacre's risk assessment documents in detail. Corporate perimeter firewalls see the device's traffic, but because the ME shares the host's MAC and IP addresses, they cannot tell ME-originated flows apart from legitimate host traffic. "The perimeter cannot attribute a flow to host-versus-CSME origin without out-of-band knowledge," Goodacre writes. A TLS-encrypted tunnel from the ME to an attacker server on port 443 looks, to the perimeter, like any other HTTPS connection the laptop makes. Network filtering reduces attack surface. It does not eliminate the exposure.
Goodacre's position is that a "Tier-3 supply-chain residual remains in both cases and is the irreducible cost of buying any silicon that ships with a Ring -3 manageability engine." He defines Tier 3 as nation state cyber services operating at the level of compromising firmware in transit, mis-issuing CA certificates via in-country authorities, and modifying hardware at customs or courier hubs. The NSA's Tailored Access Operations division treated supply chain interdiction as routine business, with explicit doctrinal preference for BIOS and firmware implants over disk-level malware.
His risk assessment's data on fleet vulnerability is concrete. Industry telemetry from Eclypsium, analyzing production enterprise environments, found that approximately 72 percent of devices observed remained vulnerable to INTEL-SA-00391 years after public disclosure, and 61 percent remained vulnerable to INTEL-SA-00295. The same reporting documented that the Conti ransomware group developed proof-of-concept Intel ME exploit code with the intent of installing highly persistent firmware-resident implants.
"Connecting an untouched-ME vPro laptop to corporate resources exposes the organization to a class of compromise that defeats the host security stack in its entirety," Goodacre concludes. "The exposed controls include BitLocker full-disk encryption, FIDO2-protected sign-in, endpoint detection and response, the host firewall and the corporate VPN."
The disagreement between Francillon and Goodacre is not about whether the vulnerability exists. Both confirm it does. Both confirm AMD faces the same issue. Both confirm software alone cannot fix it. The disagreement is about whether operational controls, Francillon's castle walls, make an architectural backdoor irrelevant in practice, or merely reduce its exploitability while leaving nation state actors with a path through.
For SecNumCloud operators processing sensitive government or commercial data, the distinction is not academic. It is worth noting that SecNumCloud is designed for a higher level of security than standard cloud certifications, but is not intended for classified or restricted government data. The threat that can still slip through Francillon's castle walls is precisely the threat SecNumCloud was designed to keep out.
The gap nobody names
Goodacre told The Register he tested awareness of the Management Engine with various attendees at the CyberUK conference in April 2026. "Almost no one" knew about it, he reports. The gap between the sovereignty rhetoric and the silicon reality is not being surfaced in policy discussions, procurement decisions, or public debate over what digital sovereignty means.
The debate that does happen, hybrid versus non-hybrid, Google/Thales versus pure European providers, focuses on operational control and legal structure. It does not address the shared silicon foundation. Strubel's LinkedIn post pushes back against the framing: "Imagining this problem is limited to hybrid cloud offerings is pure fantasy that doesn't survive confrontation with facts." Every cloud provider, hybrid or not, depends on components they don't fully control. The distinction isn't hybrid versus sovereign. It is what you're protecting against, and whether the controls you're implementing address that threat.
There is no immediate solution. RISC-V, the open source processor architecture European sovereignty advocates point to as a long-term alternative, remains years from competitive performance in datacenter workloads. "It will take decades," Francillon says flatly. Arm is a cautionary precedent: it took nearly 20 years from the first server attempts before Arm achieved any meaningful datacenter presence.
Can sovereignty exist on compromised silicon?
For Goodacre, the bottom line is simple: the Tier-3 supply chain residual is "the irreducible cost of buying silicon with a Ring -3 manageability engine." Francillon argues that operational controls, including network isolation, monitoring, and threat modeling make the backdoor unreachable except in very high-end attacks. Strubel acknowledges hardware dependencies are real but maintains that SecNumCloud provides valuable protection for what it does cover: legal control, kill-switch resistance, defense against cyberattacks and insider threats.
The disagreement is not about technical facts. It is about risk tolerance and threat model calibration. For European CIOs choosing SecNumCloud-certified providers, the question to ask vendors is: how do you address Intel Management Engine and AMD Platform Security Processor in your threat model? The answer will clarify whether the vendor treats the hardware layer as out of scope, or has implemented controls that reduce but do not eliminate the exposure.
For European policymakers, the question is broader. Can digital sovereignty exist on non-sovereign silicon? The current frameworks do not answer that question. They certify operational controls, legal structure, and autonomous execution capability. They do not certify silicon-layer immunity, because the hardware is American or Chinese, subject to American or Chinese law, designed with management engines that European authorities did not specify, cannot legally compel on their own terms, and cannot replace.
Whether that is a gap worth addressing, or a risk worth accepting as the unavoidable cost of participating in global technology supply chains, is a question Europe will need to answer for itself. ®
English (US) ·