Network Monitoring Learned the Cryptography That Is Being Replaced
In September 1752 Britain skipped eleven days. The Calendar (New Style) Act 1750 moved the country from the Julian to the Gregorian reckoning, and Wednesday the 2nd was followed by Thursday the 14th. The part worth remembering is not the astronomy. It is section 6, which provides that the Act shall not accelerate the time of payment of any rent, annuity, sum of money or interest, the delivery of any goods, or the commencement or expiration of any lease, and that nobody comes of age eleven days early. Parliament understood that correcting the frame would reach into every instrument derived from it, and it legislated for the derived instruments in the same breath as the change.
A preprint posted to arXiv on 24 August 2026 describes a small and precise version of the same frame change inside the post-quantum migration. Nobody has written its section 6.
What the paper actually measured
Bingzhen Li and eight co-authors submitted The Colossus with Feet of Clay: Debunking Encrypted Traffic Classifiers under PQC Evolution to arXiv on 24 August 2026, where it sits as a preprint and not yet as reviewed literature. Their subject is encrypted-traffic classification: machine-learning models that read the outside of an encrypted connection — packet sizes, directions, timings, burst structure — and infer what is inside without decrypting anything. Trained and tested on captures taken under the same conditions, such models score well, and the score is what gets quoted in product literature.
The authors change one thing underneath. TLS 1.3 is migrating to hybrid key establishment, and the experiments use the deployed group X25519MLKEM768, which runs the classical X25519 exchange alongside ML-KEM-768, the lattice key-encapsulation scheme NIST published as FIPS 203. Post-quantum keys and ciphertexts are much larger than elliptic-curve ones, so the handshake spreads across more packets with a different size profile. The application above is identical. The label is identical. The silhouette on the wire is different.
They call this PQC-induced protocol drift, build a benchmark pairing traditional and hybrid captures of the same activity, and run five representative classifiers and several side-channel feature representations through three settings: matched domain, cross domain, and mixed deployment ratios. The finding is sharper than a claim that monitoring breaks. The migration moves the information these classifiers depend on, and leaves it fully learnable where it lands. Feature combinations and models that perform well in-domain lose reliability once the cryptographic domain moves under them. The authors' own conclusion is methodological: matched-domain evaluation is fragile, which means the accuracy numbers this field quotes describe a condition that is currently expiring.
The scope of that result deserves to be stated as narrowly as the authors state it. It covers closed-world website fingerprinting over one deployed hybrid group, on a benchmark the authors built for the purpose. It does not measure malware detection, data-loss prevention or fraud scoring, and it is a preprint. What generalises is the mechanism: the observable form of a connection is a function of the cryptography underneath it.
Practical takeaway. Anything that reads encrypted traffic by shape — network monitoring, data-loss prevention, application visibility, anomaly baselines — carries a training baseline pegged to a handshake that is being replaced. The useful question is when it was last re-fitted against hybrid key establishment, and which team owns that answer.
Whose instrument is drifting
It is worth being exact about what the experiment studies, because the comfortable reading is wrong. The setting is closed-world website fingerprinting: recovering which site a person visited from encrypted traffic alone. That is the canonical privacy attack on encrypted browsing and the working method of network-level censorship. With different training data and a different label set, it is also the machinery an enterprise buys to see what crosses its own perimeter. One statistical technique serves the security team, the censor and the traffic-analytics vendor alike, and protocol drift does not distinguish between them.
So a migration justified by a future adversary quietly redistributes a present-day capability. A security operations team inherits a monitoring gap on a schedule it did not set. Someone browsing under a fingerprinting regime gets a real but temporary reprieve, because the observer's models are now stale and recovering them costs fresh captures and retraining. Both consequences fall out of the same experiment, and both are true at once. Neither appears in any migration plan I have read. The privacy dividend is unowned and short-lived, which is the reason to name it now, before it is found later by whoever it happens to benefit.
The mixed period is the operating condition
The deployment-ratio experiments carry the governance weight. A network in migration holds both cryptographic domains at once, in proportions that shift every month as clients, servers, middleboxes and content networks update on their own schedules. The mixed state is the operating condition, and it lasts for as long as the two domains coexist on the same wire, which makes it the state that deserves the measurement.
Those calendars stopped being speculative in the same week the paper appeared. On 24 August 2026 the US Treasury announced its Quantum-Readiness Task Force for the financial sector, following Executive Order 14412 and built on the G7 Cyber Expert Group roadmap. It runs three workstreams — sector alignment and PQC transition; third-party and vendor readiness; digital assets and emerging-technology risk — and names cryptographic agility, interoperability and third-party dependencies among its stated concerns. Migration is turning into procurement: schedules, approved products, vendor attestations, dependency inventories. That apparatus is good at asking whether a supplier supports a post-quantum algorithm. It has no field in which to record that switching the algorithm on ages a detection baseline three floors away.
A finding with no owner
The drift falls in the seam between two competent functions. Cryptographic migration belongs to platform and security engineering, measured on protocol coverage, certificate inventories and deadline compliance. Classifier baselines belong to detection engineering, measured on precision and recall against captured traffic. Each function can be complete by its own measure while the artifact between them loses accuracy. Nobody is negligent. The loss has no line to appear on.
The remedy is unglamorous and cheap while the hybrid ratio is still low: capture a hybrid baseline now, keep it beside the traditional one, and treat the pair as the thing being monitored. It costs a capture window and some retraining time. The alternative is a slow decline that produces no alert, because a model losing reliability does not announce that it has stopped recognising the traffic — it goes on returning confident labels. Where traffic inspection carries a supervisory or contractual obligation, that silence becomes a compliance question as much as an engineering one.
How Quentir Reads It
This is the third instance of one pattern in a fortnight, seen from a different floor of the building. When a journal cut the estimated quantum cost of breaking elliptic-curve cryptography, no migration deadline moved, because deadlines answer to certification calendars rather than to physics. When Microsoft dated post-quantum code signing for 2027 without naming an algorithm, the schedule again preceded the technical detail. Now a schedule set in cryptography is propagating into a system that was never part of the conversation, and someone has finally measured the effect instead of predicting it.
Traffic classification is unlikely to be the only such system. Anything trained on the observable behaviour of a protocol inherits that protocol's assumptions: capacity models keyed to session shape, congestion heuristics, fraud scoring built on connection rhythm, middlebox rules written against handshake sizes. The migration is a schema change disguised as a security upgrade, and schema changes are historically where institutions lose accuracy without noticing. That is also why this reads as a governance question more than a cryptographic one. The primitive is settled — ML-KEM is a published standard — while the deployment profiles that name it are still in motion, and the second-order consequences are landing on teams who were told the work was somebody else's.
The same question at planning scale is what the Signature Brief edition Migrated on Paper was written for: what a completed migration still leaves open once the protocol work is signed off. That edition has fixed scope, an executive summary and an internal-use license. A post like this one does the opposite — it reads a single preprint closely and stops there.
Parliament wrote section 6 because a correct change to the frame does not execute itself in everything derived from the frame; somebody has to name the derived instruments and say what happens to them. The post-quantum migration is running without that clause. The interval that matters is the one between the traffic changing shape and the models being re-fitted to it — and unlike most exposures in this field, it is measurable today by anyone willing to capture both sides of the handshake and compare. My expectation is that the first public number for that interval will come out of an incident review rather than out of a migration plan.
Sources: Bingzhen Li, Lingjia Meng, Runhan Song, Chuanzhou Pan, Tongjun Pu, Ziqiang Ma, Yupeng Jiang, Lei Cui and Zhiyu Hao, “The Colossus with Feet of Clay: Debunking Encrypted Traffic Classifiers under PQC Evolution”, arXiv:2608.22683 [cs.CR], preprint submitted 24 August 2026 and not yet peer reviewed (PQC-induced protocol drift; the deployed TLS 1.3 hybrid group X25519MLKEM768; a paired traditional and hybrid-PQC benchmark; five representative classifiers and side-channel representations evaluated under matched-domain, cross-domain and deployment-ratio settings; the finding that learnable website information persists but changes form, so in-domain performance does not carry across cryptographic domains; the recommendation that cross-domain robustness become a research priority). NIST, FIPS 203, Module-Lattice-Based Key-Encapsulation Mechanism Standard (the ML-KEM standard, of which ML-KEM-768 is the parameter set used in the hybrid group named above). U.S. Department of the Treasury, “Treasury Announces the Quantum-Readiness Task Force”, published 24 August 2026 (the public-private task force following Executive Order 14412, building on the G7 Cyber Expert Group roadmap; the three workstreams — Sector Alignment & PQC Transition, Third-Party & Vendor Readiness, and Digital Assets and Emerging Technology Risk; the stated focus on critical dependencies, cryptographic agility, interoperability and third-party implementation challenges). Calendar (New Style) Act 1750, section 6 (the saving provision: the Act does not accelerate or anticipate the time of payment of any rent, annuity, sum of money or interest, the delivery of any goods, or the commencement, expiration or determination of any lease, nor the time of attaining the age of twenty-one; 2 September 1752 was followed by 14 September 1752). All linked pages checked live on 25 August 2026.
Published intelligence, built to inform your own decisions. Published: August 25, 2026.