Post-Quantum Myths: The Call a Payments Firm Had to Make
Archi4 min read
Contents
- The Four-Month Question: Change the Keys, Not the Certificates
- Myth: This only matters once a quantum computer exists
- Myth: Post-quantum means replacing RSA and AES
- Myth: Longer RSA keys buy time
- Myth: Hybrid key agreement is a half-measure
- Myth: It is a configuration flag at the edge
- Myth: Compliance will tell you when
- What They Shipped
- The Part Nobody Plans For
The Four-Month Question: Change the Keys, Not the Certificates#
The team had four months and one narrow question: could they change how TLS sessions agree on a key without touching a single certificate? The company processes card-not-present payments, keeps transaction records for seven years, and terminates TLS on a pair of load balancers whose firmware had not been updated since 2023. That combination set the deadline. Here is what the myths looked like from inside the room.
Myth: This only matters once a quantum computer exists#
Fact: the threat model is recording, not breaking. An adversary who copies encrypted traffic today can hold it until decryption gets cheap. For a company with a seven-year retention window, the exposure period runs from the day a session was captured to the day the data stops mattering. Add two years to migrate and the arithmetic gets uncomfortable fast. The arrival date of the hardware is a detail. The retention policy is the clock.
Myth: Post-quantum means replacing RSA and AES#
Fact: ML-KEM, standardized by NIST in FIPS 203 in August 2024, replaces the key agreement step. It does not replace AES, and AES-256 stays usable: Grover's algorithm roughly halves the effective security level, which still leaves a margin nobody is attacking. Certificates and signatures are a separate migration on a slower timeline. Mixing the two up is how teams spend a year on the wrong problem.
Myth: Longer RSA keys buy time#
Fact: RSA-4096 is not a hedge against Shor's algorithm, it is a bigger number in the same broken shape. It also costs handshake CPU and bytes every connection. The company's security lead put it bluntly in the working group: adding key length to the handshake was the one option that increased cost and reduced nothing.
Myth: Hybrid key agreement is a half-measure#
Fact: hybrid means using both. The X25519MLKEM768 group combines a classical elliptic-curve exchange with ML-KEM-768, so an attacker has to break both to recover the session key. If ML-KEM has a flaw, X25519 still holds; if a quantum machine arrives, X25519 stops mattering but ML-KEM does not. Chrome and Firefox began negotiating hybrid groups by default in late 2024, which meant the company's own browsers would soon be testing the path whether the backend was ready or not.
Myth: It is a configuration flag at the edge#
Fact: the flag is the easy part. The real work was inventory. Which terminators could negotiate the new group at all? Did the session ticket encryption keys need rotating? Could the pinned mobile TLS stacks on older Android builds handle a larger ClientHello? And what about the HSM sitting behind the payment flow, whose firmware had no ML-KEM support and no committed date for it?
Size was the sleeper issue. ML-KEM-768 carries a 1,184-byte encapsulation key, so the hybrid ClientHello grows by roughly 1.2 KB and the server response by about 1.1 KB. On a clean network that is nothing. On a lossy mobile link it can push the first flight past the initial congestion window, adding a round trip before the handshake even starts. The team found handshake failures clustered in exactly the markets where their mobile traffic was weakest.
Myth: Compliance will tell you when#
Fact: guidance is not a schedule. NIST's draft IR 8547 proposes deprecating RSA-2048 and P-256 after 2030 and disallowing them after 2035, and those are draft dates for federal systems, not a mandate for a payments processor. The company's own lawyers asked a better question: what does our data retention promise require us to protect in 2036? That answer set the timeline, not a slide from a conference.
What They Shipped#
The company enabled the hybrid group at the edge first, kept classical groups as fallback, and tuned session resumption so repeat connections paid the size cost once rather than every request. Rollout went 5 percent, then 25, then everything, with a dashboard watching handshake failure rate by region and device class. Certificates stayed classical, deliberately, with a written note explaining that signature migration was a separate project with a separate owner and date.
They also built a list of every vendor in the request path that could not negotiate the new group, and put a date next to each name. Two vendors got a year. One got a formal exception from the security lead, signed and filed.
The Part Nobody Plans For#
Latency held. ML-KEM is fast enough that the CPU cost never showed up in the graphs; the bytes did. The surprise was operational: a small percentage of clients silently fell back to classical groups, and without per-group telemetry the team would have believed coverage was complete when it was not. The lesson they passed on was unglamorous. Measure which group actually won each handshake, keep the classical fallback until the failure data says otherwise, and treat the signature migration as a different job with a different deadline.
More articles