Cybersecurity research · Tokyo

ba-cybersecurity.com

Reporting recent research activity by Nobuo Otoi.

As of · 19 Nov 2026Next · IEICE IA, 3 Dec 2026

Show

Recent

19 Nov 2026 · NSA cipher ARADI

The NSA cipher ARADI: the six-round optimum is 82, not the 84 printed as optimal, and the nine-round linear value 71 is optimal

Table 4 of CLAASPing ARADI prints 84 in bold as the optimal six-round differential weight; inside CLAASP's own model the optimum is 82, shown by a weight-82 trail and 945 subproblems with independently checked DRAT proofs. The same method confirms the eight-round values 122 (differential) and 62 (linear), and shows that the nine-round linear value 71, printed without a claim, is optimal. Not an attack.

Preprint v3, 19 Nov 2026More →

Web application defense

Predicting a configuration change before deploying itIn preparation

Whether a change to OWASP CRS settings will cause false positives or lose detection, known before it goes live. Prospective validation is in preparation.

Verifying Internet standards

ECDSA scalar rule (CA/B Forum draft)Comment on the ballot

A proposed MUST — reject ECDSA keys whose private scalar is below 256 — accepts keys that are just as cheap to recover: n − k (same x-coordinate as k·G) and k⁻¹ mod n. Shown on the eight eMudhra keys, with a one-sentence correction and one equivalence proved in Lean 4. The negation family is already in zlint's lint; the ballot text and the inverse family are not. The ballot author agreed these keys are just as easy to recover, while arguing the Baseline Requirements should not require checking for them.

k = 4 (an eMudhra scalar), d = n − 4 draft d < 256 ? no → accepted same x as 4·G → recovered in ≤255 steps
ECDSA scalar: check it by hand
The sentence (PR #683, §6.1.6): reject keys whose private scalar is less than 256. Take k = 4 and d = n − 4 (n the P-384 order). d is not less than 256, so it is accepted. Its public key (n−4)·G = −(4·G): same x-coordinate as 4·G. The eMudhra search, compare x(Q) with x(k·G) for k = 1..255, returns k = 4. Second family d = 4⁻¹ mod n: then 4·Q = G, found in four additions of Q. one sentence, two families it does not name fix: neither d nor n − d < 2^16, and no k < 2^16 with k·Q = ±G

CA/B servercert PR #683 · Bugzilla 2083449

CRI (IETF CoRE draft)With the authors

A CRI reference and its URI form could resolve to different addresses. 3 mismatches with RFC 3986 were shown with counterexamples and fixes, one fix proved in Lean 4. An author said they will be corrected before publication.

[1, ["", "a"]] against s://h/b/c §5.3 s://h/b//a §6.1 + 3986 s://h/a
CRI: check it by hand
Base s://h/b/c, reference [1, ["", "a"]] §5.3 resolves the reference s://h/b//a §6.1 converts it to the URI reference /a, and RFC 3986 resolves that s://h/a One reference, two different results.

Draft §5.3 · Draft §6.1

ASRA (IETF SIDROPS draft)With the authors

BGP path verification could mark a correct route Invalid. This was shown with a counterexample and a fix proved in Lean 4. An author said the algorithm will be revised.

correct route AS0, AS3, AS2 -05 Invalid fix not Invalid
ASRA: check it by hand
Route AS0 → AS3 → AS2 → AS1 Up to a provider, across to a peer, down to a customer: no leak. Only AS3 publishes: an ASPA, and an ASRA listing AS2 as a lateral peer. AS0 and AS2 have no ASPA, so min_up_ramp = 1, min_down_ramp = 1, N = 3. Step 4 checks i from (min_up_ramp + 1) to (N - min_down_ramp), so i = 2. AS(2) = AS3 asserts that AS(3) = AS2 is a lateral peer. -05 Invalid fix start at max_up_ramp + 1 = 3: nothing to check, not Invalid

Draft §4.2, Algorithm A

SUIT (IETF SUIT draft)Fixed in -17

A firmware condition for a final release also accepted its prereleases. This was shown with a counterexample checkable by hand. An author confirmed the error, and revision -17 of the draft (2 Nov 2026) changed the encoding; an exhaustive check and a Lean 4 proof support the fix.

Equal [1,2,3] against 1.2.3-alpha.4 draft matches SemVer lower than 1.2.3
SUIT: check it by hand
§4.4.1 1.2.3-alpha.4 = [1,2,3,-3,4] §4.4 comparison "stops after all integers in the array defined by the manifest have been consumed" Equal [1,2,3] against 1.2.3-alpha.4 1 = 1, 2 = 2, 3 = 3, array consumed, stop draft matches SemVer §11 1.2.3-alpha.4 is lower than 1.2.3

Draft §4.4 · SemVer §11

Japanese municipal websites: hosting and RPKIDataset

Where the official websites of all 1,741 Japanese municipalities are hosted, and whether the route to each is protected by RPKI, measured passively on one day in October 2026 and released with a rebuild script.

Security review of widely used software

OAuth servers and the RFC 10070 audience rulePublished

RFC 10070 forbids an OAuth server from accepting its token-endpoint URL as a client-authentication audience; it must be the issuer identifier alone. Five major open-source servers were read at pinned commits — all five still accept the token-endpoint URL. A conformance snapshot, Low severity, not a vulnerability.

FreeRADIUS (network authentication server)Fixed

TEAP could accept a client without verifying any MAC. Reproduced on 3.2.10 and reported; the developers fixed it and credited the report.

Langfuse (LLM engineering platform)Fixed

Any organization member could invite users, without the required role, by naming a project from another organization. Reported; the developers fixed it and listed the report in their acknowledgements.

python-cwt (COSE cryptography library)Fixed

A signed or encrypted COSE message whose protected header repeated the same label (for example, the algorithm given twice) was accepted, silently keeping the last value, where RFC 9052 requires it to be rejected. Reported; the maintainer fixed it the same day and released it in 4.0.0.

Mealie (self-hosted recipe manager)Fixed

Any user could move a notifier into another household and receive that household's recipe, shopping-list and meal-plan activity. Reported; the maintainer fixed it in 3.31.0 and published advisory GHSA-qwhr-6537-wj7f (Moderate), crediting the report.

http-message-signatures (npm, RFC 9421) — signature verification bypassFixed

Signature verification bypass (sender impersonation). In HTTP request signing (RFC 9421), the npm library could be made to accept a forged signature: with all: true ("verify every signature"), it checked only the last signature it had a key for. A sender holding only key B could therefore pass a message as signed by key A too. Applications that trust every listed key ID could be fooled; none of four public users checked was affected. Reported with a fix and tests; the maintainer extended it and released it in 1.0.7.

wallabag (self-hosted read-it-later app)Fixed

Any logged-in user could attach annotations to another user's saved entries: the create path skipped the ownership check that the read, update and delete paths enforce. Found independently and reported, as it was by several other researchers; the maintainers fixed it in 2.6.16 and published advisory GHSA-q2r9-jqp3-5f26 (Medium), crediting this report among twelve reporters.

Credo (digital-credential wallet and verifier framework) — two verification flawsFixed

Forged or unsigned data could pass verification. Credo is the open-source framework (OpenWallet Foundation) that wallet and verifier apps use to check digital ID cards, licences and certificates. A holder could hand over one genuine credential plus a made-up block of details, and Credo returned the made-up block as the checked credential. Separately, a did:webvh public record (schema or credential definition) in an issuer's name was accepted with a valid signature from someone else (rated theoretical by the maintainers: it also needs control of where the issuer's records are served). Both reported privately; the maintainers fixed them in 0.7.5 and published two advisories (Moderate) listing this report.

net-snmp (npm package for Node.js, not the Net-SNMP C tools) — authorization bypassFixed

A blank user name bypassed authorization. SNMP is how network devices and servers are monitored and configured. In Node.js SNMP agents built on net-snmp with default settings, an unauthenticated request with an empty user name could read every value served, change writable ones, and inject alerts into a receiver. Reported privately with a suggested fix; the maintainer adopted that approach, the researcher tested the fix branch, and 3.29.3 was released within a day with advisory GHSA-pg9p-672x-44jc (Critical, 9.1) crediting him as reporter.

Other projectsWith the developers

Problems have been reported to several projects and are being reviewed. They will be added here once fixed.

Open values in symmetric cryptography

Ascon: fewer active S-boxesPreprint

Lowered the best published active S-box counts we could find for NIST's lightweight standard. One pair of states follows all four rounds, and anyone can check it in a minute; 5 rounds have only a per-round check. Neither threatens the security of Ascon.

Ascon: check it yourself
$ python3 verify/check.py round 8: 15 active S-boxes round 9: 11 active S-boxes round 10: 9 active S-boxes round 11: 7 active S-boxes total: 42 matches the characteristic: True Best published value we could find: 43. One pair of states, no dependencies, about a minute.

verify/check.py

SPECK128 (NSA), 12 roundsPreprint

Weight 66 is optimal: no 12-round characteristic of the NSA lightweight cipher SPECK128 has weight 65 or less. The best published value was 66, with optimality open beyond 11 rounds (IEICE 2025). The 272 SAT subproblems are all unsatisfiable; the split and the 89 shortest subproblems carry DRAT certificates, and a second, separately written encoding solved by a different solver gives the same answer. Not an attack.

No solution: 36

ARADI (NSA), 6 to 9 roundsPreprint

Table 4 of CLAASPing ARADI prints 84 as the optimal six-round differential weight of the NSA cipher ARADI; inside CLAASP's own model the optimum is 82, with DRAT-checked proofs that nothing lighter exists. The same method confirms the eight-round values 122 (differential) and 62 (linear) and shows that the nine-round linear value 71 is optimal. Not an attack.