SEALSQ Research Exposes Side-Channel Risk in ML-KEM
A chip vendor has published research on an attack against the post-quantum standard its hardware is designed to protect. This news emerged on October 14, 2026, when SEALSQ presented a peer-reviewed paper at LightSec 2026, a workshop affiliated with CHES in Antalya.
The paper describes a side-channel attack targeting the encoding step of ML-KEM key generation. ML-KEM, also known as NIST FIPS 203, is a lattice-based key-establishment scheme formerly called CRYSTALS-Kyber. SEALSQ, a Swiss semiconductor company, sells post-quantum secure elements, including its Quantum Shield QS7001 chip, which directly embeds ML-KEM in silicon. It’s important to note this upfront: the company that discovered this vulnerability also offers the countermeasure.
This raises the question of how this finding impacts the trustworthiness of FIPS 203. Let’s examine the strongest potential concern and then evaluate its validity.
The Strongest Interpretation of the Claim
The most severe interpretation is as follows: if practical attacks can extract ML-KEM private keys from operational hardware, then the migration efforts currently underway by NIST, Europol, and a significant portion of the crypto industry are built on a flawed foundation. This concern isn’t theoretical. Starknet already deploys a quantum-resistant account path using post-quantum signatures. Ethereum researcher Justin Drake has warned that wallet security could be compromised by advancements in AI and mathematics even before quantum computers materialize. Real-world systems are relying on these algorithms now, so a reliable implementation attack would therefore be significant for all of them.
To assess this claim, we can break it down into several parts:
- The attack recovers ML-KEM secret-key information, not just random noise.
- It operates on physical hardware under realistic conditions.
- It targets the standardized algorithm broadly.
- The leakage is severe enough to necessitate a fundamental reevaluation of FIPS 203 itself.
Each of these points can be independently true or false. Our audit will categorize them accordingly.
What the Evidence Shows
Let’s compare the strong claim with the paper’s findings. The paper, titled “Cross-information lattice attack on ML-KEM,” combines profiling-based side-channel techniques with lattice reduction to recover sensitive information during the encoding step of key generation. According to SEALSQ’s own report, the results hold “under realistic conditions,” and the paper includes proposed countermeasures.
Now, for precise categorization: The attack targets an implementation, not the underlying mathematical principles. Side-channel leakage arises from power consumption, electromagnetic emissions, or timing, none of which undermine the hardness assumption that underpins the security of lattice schemes. The research was presented at a CHES-affiliated workshop and described as peer-reviewed, which is more credible than a self-published blog post but less conclusive than a finding already independently reproduced by the wider community.
SEALSQ states that their method recovers information “more effectively than prior methods,” positioning it as an incremental advancement within existing academic work on Kyber and ML-KEM side channels. Furthermore, the paper offers defenses, indicating that the authors view this as an effort to strengthen, not dismantle, the standard.
What the available material doesn’t clarify is the question of full recovery. The press release doesn’t state whether the attack recovers a complete private key or merely provides leakage statistics on partial data. It also doesn’t specify whether the attack was performed on SEALSQ’s own silicon, a reference implementation, or an open-source library. This crucial omission directly impacts the overall assessment, so we must reserve judgment here.
Beneficiaries and Data Origins
A vendor selling hardened post-quantum chips publishing attack research against the very standard those chips implement inherently creates demand for its specific countermeasure hardware. The QS7001 advertises dedicated defenses against SPA and DPA attacks, active shielding, and scrambling, and is currently undergoing Common Criteria EAL5+ certification.
Both the LightSec paper and the product line address the same class of risk. CEO Carlos Moreira characterized the work as evidence of the company’s “dual role as both product innovators and contributors to the global cryptographic research community.”
The incentive structure doesn’t automatically invalidate the research. Peer review at a CHES-affiliated venue provides a genuine filter, and side-channel attacks on lattice encoding are a legitimate area of research that predates this paper. The caution is more specific: the result is currently self-published via a press release and conference presentation; no independent group has yet replicated it based on the available information; and SEALSQ has not disclosed whether the target implementation was its own, a competitor’s, or a generic one. This last omission influences how one interprets the messenger. It’s currently unclear whether the company attacked its rivals or stress-tested its own technology.
Implications for ML-KEM
The evidence suggests this is a matter of implementation hardening, not a broken standard. If the demonstrated attack is an instance of implementation leakage under controlled conditions against a specific build, then FIPS 203 remains widely considered secure, and the primary risk shifts from the algorithm itself to engineering practices.
RSA and ECC underwent a similar evolution for decades, where timing and power attacks led to improved implementations without invalidating the underlying mathematics. Side-channel exposure is a known and persistent concern for lattice schemes, particularly on resource-constrained and long-lived devices, precisely where secure elements are most valuable.
The deployment context elevates the stakes but doesn’t alter the fundamental assessment. Starknet’s production use of Falcon-based accounts and ML-DSA-65 signatures means that implementation quality is already paramount. Europol has identified wallets as the primary vulnerability point while advocating for a phased migration. Drake’s warning regarding ECDSA demonstrates that urgent claims can circulate with varying degrees of supporting evidence. None of this suggests ML-KEM is inherently weak. Rather, it indicates a wide audience for these findings.
Factors that would shift the verdict:
- Independent replication of full key recovery by a group with no commercial interest.
- A demonstrated attack on deployed hardware rather than a laboratory setup.
- Comments from NIST, the IETF, or the Kyber team indicating this is a standard-level concern.
- Disclosure that the target was certified, protected silicon rather than an unprotected reference build.
Absent these, the responsible conclusion is that ML-KEM remains believed secure, the finding represents valuable implementation research, and a vendor’s financial interest in the conclusion warrants careful reading, not outright dismissal.