Daily Summary for 2026-10-25

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: It's the twenty-fifth of October, twenty twenty-six, and this is the day's research.

Elias: 14 new papers came out today.

Nadia: I'm Nadia.

Elias: And I'm Elias, here with Priya.

Nadia: We'll take the day in one pass, then pull out the papers we're staying with.

Elias: Start us off with When EVM Defaults Break.

The day's research: Nadia: Today's work focuses on adapting AI models to improve performance in real-world security tasks.

Elias: That moves beyond static models to something more responsive.

Priya: We looked at steering emotion directions within the AKMPC transformer.

Nadia: Nudging the model toward calm or desperate states improved performance by one and a half to two percentage points.

Elias: So emotion representations are controllable tools for adaptive reasoning, is that right?

Priya: Another area is how large reasoning models expose private location data through visual clues.

Nadia: The GeoTrace attack showed models like Gemini-3-Flash localize ordinary images within one kilometer about twenty-six percent of the time.

Elias: That localization relies heavily on small, distinctive local clues instead of broad scene properties.

Priya: To counter this, the GeoShift defense preserves stable scene features and rewrites revealing clues to suggest an alternative location.

Nadia: GeoShift successfully reduced street-level localization significantly because of that rewriting process.

Elias: We also examined how multimodal models can be tricked into bypassing safety checks, like in Mirage.

Priya: Mirage demonstrated that imperceptible perturbations can align an image with a malicious target while evading post-generation safety checkers.

Nadia: That means the model bypasses checks across different diffusion models using subtle changes.

Elias: On the resource side, TokenTrap disrupts LLM-based scam systems by redirecting fraud attempts toward low-value computations.

Priya: It amplifies token consumption dramatically across different agent workflows to achieve that redirection.

Nadia: Data voids in web searching LLMs mean disinformers succeed even with one controlled page because the model relies on sparse information when answering data-void prompts.

Elias: While user prompts offer slight protection, attacker content appears in responses thirty-one percent of the time under various conditions, highlighting a need for defenses against this inference-time manipulation.

Priya: How can we make large language models useful where resources are limited?

Nadia: This is addressed by work on parameters efficient language model with latent cross-attention for IoT intrusion detection.

Elias: That matters because it shows we can create a model small enough to run on tiny devices while still being accurate enough for security tasks.

Priya: Does this approach compress the input into a small latent space before using a minimal Transformer core?

Nadia: Yes, and comparing nine compression configurations against SecurityBERT on two datasets showed their model is about seven times smaller but achieves similar or better performance in intrusion detection metrics.

Elias: So lightweight Transformer architectures can be practical for security monitoring on resource-constrained IoT devices.

Nadia: PhishGen measures LLM capability by generating pages against sixty-four real-world kits.

Elias: How does that measure relate to the actual credential harvesting pipeline completion rate?

Nadia: Only a fraction of outputs complete the full credential harvesting pipeline, with most outputs rendering.

Priya: HiMark tracks text origin using hierarchical mechanisms during embedding and model-agnostic evidence aggregation for detection.

Elias: Does that framework maintain text quality when paraphrasing attacks are applied?

Priya: Yes, it improves watermark extraction accuracy while preserving text quality under paraphrasing attacks.

Nadia: Who Patches the Lightning Network shows upgrades happen based on platform releases, not security disclosures.

Elias: That contrasts with expectations about when patches are actually deployed in the network.

Priya: The network patches on release and platform schedules, which covers a small window of time for upgrades.

Nadia: When EVM Defaults Break reveals three SDK patterns for UTXO-family wallets to guide design.

Elias: What are those three recurring patterns extracted from the retrospective case study?

Nadia: There is a Transaction Composer, a Unified Signer Abstraction, and a Connector UI component.

Priya: The contribution is articulating these as transferable design lessons for SDKs targeting UTXO-family ledgers.

Elias: BelT Block Cipher describes a block cipher with 128-bit blocks and 256-bit keys.

Nadia: It employs an eight-round structure combining elements of a Feistel network and the Lai-Massey scheme.

Priya: Its specification was issued as a preliminary standard in 2007, becoming a final standard in 2011.

Nadia: We also reviewed research on emotion representations to improve cybersecurity AI performance.

Elias: What specific findings were observed regarding steering toward calm or desperate emotions?

Nadia: Steering toward calm, serene, or desperate improves performance by 1.6 to 2.1 percentage points across ten seeds.

Priya: Results depend strongly on the steering direction, layer, and intensity of the emotion applied.

Elias: Not Where You Think explores how reasoning models infer location from visual information in photos.

Nadia: GeoTrace exposes how models extract location-relevant visual clues through coarse-to-fine geographic reasoning.

Priya: Broad scene properties support coarse localization, while precise localization relies on small, distinctive local clues.

Elias: GeoShift is a proactive defense that rewrites high-impact geo-revealing clues to create an alternative hypothesis.

Nadia: This defense reduces street-level localization from 25.71% to 3.81% under Gemini-3-Flash.

Priya: Mirage shows that post-generation safety checkers can be bypassed by a multimodal jailbreak attack called Mirage.

Elias: How does it exploit the shared representation space between image and text encoders?

Priya: It exploits this space to align the perturbed image with the target while repelling filtered concepts.

Nadia: Your Token Is Gone presents TokenTrap, a defense that redirects scam systems toward benign, low-value computation.

Elias: How does this countermeasure amplify token consumption in the observed workflows?

Nadia: TokenTrap amplifies token consumption by 2, 487.0 times across the three workflows tested.

Priya: Rank Ceiling for Twiddle-Perturbation Faults on ML-KEM and ML-DSA characterizes exposure during NTT computation faults.

Elias: For one fault, rank equals the number of butterflies sharing its twiddle, halving by layer from 128 to 2 for ML-KEM.

Nadia: Membership Inference Attack on Private Evolution introduces an attack tailored to the voting-based algorithm in PE.

Priya: This attack achieves empirical performance indicating private data leakage across image, text, and tabular modalities.

Elias: Data-Void Exploits in Web-Searching LLMs study how disinformers use sparse content voids for narrative manipulation.

Nadia: Attacker-content appears in LLM responses 31% of the time across all tested conditions.

Priya: HAWC-Leak uses a dynamic workload classification framework to identify hardware accelerator tasks via power analysis.

Elias: Does this framework achieve 100% classification accuracy across all evaluated mixed-activity scenarios?

Priya: Yes, it confirms that power consumption patterns serve as a reliable fingerprint for profiling runtime behavior.

Nadia: Parameters efficient language model with latent cross-attention proposes a compact architecture for IoT intrusion detection.

Elias: With only 306k non-embedding parameters, how does this model compare to SecurityBERT?

Priya: It outperforms SecurityBERT by about 2 points of macro F1-score on Edge-IIoTset.

Nadia: What about the research on LLM capabilities in automated phishing website generation?

Elias: PhishGen assesses functionality against 64 real-world phishing kits using varied LLMs.

Priya: Only 38.7% complete the full credential-harvesting pipeline, showing a trade-off for scalable generation.

Nadia: The work Who Patches the Lightning Network provides insight into how network maintenance occurs in practice.

Lucky paper: 2610.30183: Nadia: So we're diving into When EVM Defaults Break, looking at how SDK design changes when you move to extended-UTXO models.

Elias: That paper looks at how mainstream Web3 SDKs, like ethers.js and viem, make three assumptions that fail once a chain adopts an extended-UTXO model or uses different cryptographic primitives like Schnorr keys.

Priya: The paper shows that the extended-UTXO setting imposes three qualitative differences on walletside SDK design compared to the EVM case.

Nadia: That first difference is that every UTXO-like unit, or a cell, must carry enough capacity to cover its serialized output and data, so transaction construction has to balance inputs and outputs against capacity, not just value.

Elias: That second difference is that the chain is script-pluralistic; a lock script can encode any signature scheme, meaning different unit owners need different signing pipelines even inside one transaction.

Priya: And the third difference is that fee estimation becomes coupled to transaction byte size and lock-script choice, creating a feedback loop that isn't there in the EVM gas model.

Nadia: The CKBers’ Codebase, or CCC, extracts three design patterns from its release history to address these constraints.

Elias: Pattern A is the Transaction Composer; it means the caller supplies to/value/data, and the SDK fills in nonce and gas, resolving inputs, change units, byte size, and fee in one pass.

Priya: That resolves the coupling between inputs and fee estimation by accepting declarative outputs where the SDK resolves everything at once.

Nadia: Pattern B is the Unified Signer Abstraction; it normalizes signing across different schemes because secp256k1 ECDSA, WebAuthn passkeys, and Schnorr keys can coexist on one chain.

Elias: So that means one signer interface encapsulates all those per-scheme differences and drives eight heterogeneous wallet adapters.

Priya: Pattern C is the Connector UI; this handles per-wallet detection, chain switching, and account-change handling by shipping a drop-in web component or React wrapper around the Signer interface.

Nadia: The paper concludes these patterns generalize to other extended-UTXO ecosystems, like Cardano, suggesting declarative composition and reusable connectors are already appearing.

Elias: But the most distinctive contribution is Pattern B’s unification across signing primitives because it lifts wallet abstraction out of the application layer.

Priya: The central lesson is that an SDK in this regime should hide routine complexity by default while keeping the manual path open, and they test whether these patterns reduce integration time or defects.

Lucky paper: 2610.31178: Nadia: Okay, let's look at BelT Block Cipher now. The main thing here is that it’s a block cipher with a one hundred twenty-eight-bit block size and a two hundred fifty-six-bit key that was adopted as part of the Belarusian national standard STB thirty-four point one zero one.thirty-one.

Elias: That's significant because it shows how established standards are being used for actual encryption and integrity control. What about the structure?

Priya: The design uses an eight-round structure that combines elements of a Feistel network and the Lai–Massey scheme for both encryption and decryption.

Nadia: And it starts by initializing three thirty-two-bit registers as a, b, c, d, and then the encryption involves rounds one through eight while decryption performs rounds eight through one.

Elias: That structure is what simplifies the software and hardware implementations, which is a design principle mentioned in the paper. How does that eight-round structure relate to security?

Priya: Preliminary statistical analysis showed that no significant statistical regularities were present for a two-round construction, which motivated the choice of eight rounds to establish a comfortable security margin.

Nadia: The key schedule generates round keys kappa i from a two hundred fifty-six-bit key theta, where kappa i is the i-th element of the sequence theta1, theta2,, theta8.

Elias: That means they have to calculate these round keys before they can even start the encryption rounds. What about the transformation function G r ?

Priya: The transformation Gr maps a thirty-two-bit string u = u1 u2 u3 u4 to Gr(u) = lambda r(S(u1)S(u2)S(u3)S(u4)).

Nadia: The S-box, denoted as S, maps A8 to A8, and its properties are based on truth tables of coordinate Boolean functions selected as distinct length-two hundred fifty-five segments of linear recurrence sequences.

Elias: I see the focus there on the nonlinearity of S being one hundred two which provides resistance to linear cryptanalysis. What about the differential properties?

Priya: The differential properties are characterized by quantities like R (S) = eight R (S) = six and R (S) = three.

Nadia: Those values are sufficiently small to provide resistance to differential cryptanalysis. Plus, the algebraic degree of every coordinate Boolean function of S is seven.

Elias: A degree of seven is important because it hinders higher-order differential attacks, which is a point you brought up earlier about the design principles.

Priya: Exactly, there are no quadratic relations connecting the inputs and outputs of the S-box, which simplifies algebraic attacks.

Nadia: The paper also looked at performance, showing that a thirty-two-bit architecture was chosen as the basis for this cipher's design.

Elias: How does that performance compare when you look at benchmarking results against candidates in the Advanced Encryption Standard competition?

Priya: BelT achieved an encryption speed of two hundred times one hundred twenty-eight times one thousand three hundred sixty-seven about eighteen point seven three Mbit/s on a Pentium-two hundred CPU.

Nadia: That speed is compared against AES, where BelT achieves a speed of six hundred forty-two cycles on a Pentium II CPU.

Elias: So the overall design aims to balance strong cryptographic properties with practical computational performance for that thirty-two-bit architecture.