New models and updates from major AI providers this week
NVIDIA Nemotron 3 Ultra has been integrated into CrowdStrike's SafeMind agentic cybersecurity system to orchestrate defensive agent harnesses. The model serves as a core component in an adversarial coevolution loop designed to automate cyber defense.
OpenAI has released the safety overview and system card for GPT-6 Astra, detailing its security protocols. This release focuses on establishing frontier safeguards and critical capabilities for this next-generation model.
StepFun has released Step 3.7 Flash, a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model designed for high-frequency production workloads and agentic workflows. The model features native image understanding with an 11B active parameter count per token and supports up to 256k context window.
Anthropic's Claude Fable 5.1 has been released on Amazon Bedrock and the Claude Platform on AWS. The release focuses on enhanced model performance alongside new Enterprise Frontier Safeguards to ensure data security within controlled cloud environments.
Google DeepMind has introduced agentic video understanding capabilities within the Gemini model. This update focuses on enhancing how models process and reason through video content using an agentic approach.
AWS has highlighted its Nova foundation models which are designed to deliver frontier intelligence alongside industry-leading price performance. These models serve as a core component of AWS's generative AI offerings, providing high-scale capabilities for various workloads.
Selected arXiv and HuggingFace papers this week
Paper 1
The paper introduces SecOPD, a new defensive fine-tuning method using token-level on-policy distillation to mitigate adaptive prompt injection attacks in LLMs.
TL;DR
Researchers have developed SecOPD, a fine-tuning technique that uses token-level feedback to protect AI agents from prompt injection attacks. This method significantly outperforms previous state-of-the-art defenses by precisely identifying and penalizing malicious tokens during training.
Prompt injection remains a critical vulnerability for AI agents that interact with external data sources like emails and websites. Current defensive fine-tuning methods, such as Direct Preference Optimization (DPO), rely on sequence-level feedback, which treats an entire response as either good or bad, thereby obscuring the specific tokens responsible for security breaches. To address this, the authors propose Secure On-Policy Distillation (SecOPD). This approach utilizes token-level feedback by using a 'teacher' model—an instance of the base LLM run on clean, injection-free data—to score individual tokens produced during the student model's rollout. By providing more granular supervision, SecOPD allows the model to learn exactly which parts of a hybrid response (partially benign, partially malicious) should be ignored. Experimental results show that when applied to Qwen3.6-27B, SecOPD reduced the Attack Success Rate (ASR) against the PISmith adaptive attack from 94% to just 9%. Furthermore, the method proved effective in unseen scenarios like agentic tool calling and maintained the model's original performance utility.
Paper 2
The introduction of CyberFactory, an open-source framework designed to scale cybersecurity capabilities in LLMs by transforming real-world CVE data into verifiable training trajectories.
TL;DR
Researchers have developed CyberFactory, a unified framework that automates the creation of high-quality cybersecurity training data from real-world vulnerabilities. By using agentic trajectories to train the OpenAegis model, they achieved state-of-the-art performance in vulnerability detection and patching.
The research addresses the gap in open-source cybersecurity LLM development, where existing models lack reproducible training methods and scalable agentic data. The authors propose CyberFactory, an end-to-end pipeline that transforms public vulnerability artifacts, such as CVEs, into executable task instances for proof-of-concept generation, patch generation, and cybersecurity question answering (CyberQA). A core innovation of the framework is its use of 'agentic supervision,' where a teacher model utilizes a specialized vulnerability-analysis skill to perform source inspection, problem-solving via domain priors, and evidence-based validation. This process involves interacting with target environments and revising solutions based on execution feedback, creating highly reliable training trajectories. Using this synthesized data, the researchers trained OpenAegis, an open-weight model that successfully internalized the complex analytical workflows of its teacher. In evaluations conducted on the CyberGym benchmark under a one-hour budget, OpenAegis achieved a 58.1% Pass@1 rate, representing a massive 28.5 point improvement over its Qwen 3.5 base and outperforming much larger models like GLM 5.2 and Kimi K2.7. The study concludes that verifiable, tool-augmented trajectories can effectively teach models both specific task outcomes and coherent, systematic security analysis procedures.
Paper 3
The introduction of Test-Time Policy Optimization (TTPO), a new method for improving large language model reasoning during test-time training without ground-truth labels.
TL;DR
The paper introduces TTPO, a method designed to optimize LLM reasoning during test-time training by utilizing asymmetric learning signals from pseudo-labels. It effectively mitigates the risks of noisy majority-vote labels by distilling agreeing rollouts and penalizing disagreeing ones.
The research addresses a critical limitation in current post-training methods for Large Language Models (LLMs), such as Reinforcement Learning (RL) and On-Policy Self-Distillation (OPSD), which rely on ground-truth labels. In test-time training (TTT) scenarios where labels are unavailable, researchers often resort to majority-vote pseudo-labels; however, these labels are frequently incorrect in complex competition-level tasks, leading to corrupted teachers and misleading updates. The authors identify an asymmetric failure mode: while a wrong pseudo-label can mislead distillation, rollouts that disagree with the pseudo-label are statistically likely to be incorrect regardless of the label's accuracy. To exploit this, the authors propose Test-Time Policy Optimization (TTPO). TTPO applies OPSD distillation specifically to rollouts that agree with the pseudo-label and utilizes Grouped RL (GRPO) to penalize rollouts that disagree with it. This approach is further refined by token-level selection, which focuses updates on unmastered positions and confident errors. Experimental results show that TTPO, trained entirely without labels, matches or exceeds label-supervised OPSD across various Qwen3 scales. Specifically, for the Qwen3-1.7B model, accuracy increased from 38.0% to 45.2% in TTT settings, with even larger gains observed when 'thinking mode' is disabled. The method also demonstrates robust cross-task generalization, suggesting that the learned reasoning capabilities are not merely overfitting to specific benchmarks.
Paper 4
The paper investigates the fundamental unreliability of using black-box LLM judges for evaluation due to non-deterministic outputs caused by shared endpoint engineering.
TL;DR
This research highlights the critical failure of using black-box LLMs as evaluators due to inherent non-determinism in shared computing environments. It advocates for a rigorous, preregistered auditing framework to ensure measurement stability and reproducibility in AI evaluation.
The paper addresses the growing reliance on 'LLM-as-a-judge' paradigms, such as MT-Bench and Chatbot Arena, by exposing a fundamental flaw: the lack of reliability in black-box API-hosted evaluators. While previous literature has documented biases like position bias, verbosity, and self-preference, this study focuses on the technical instability caused by the underlying engineering of shared endpoints. Specifically, the authors demonstrate that because kernels are not batch-size-invariant and greedy decoding can diverge due to drifting logits during concurrent request batching, even identical inputs can yield different results. The researchers implement a novel 'instrument-audit' setting, where they use preregistered gates to determine if an evaluation program should be terminated based on measured inconsistency. Unlike existing reproducibility checklists that focus on hyperparameters or code, this work emphasizes the need for snapshot-identity level tracking and measuring the noise floor of the instrument itself. By applying principles from analytical chemistry—such as detection limits and gauge R&R—the authors provide a framework for decomposing instability into bias, separation, and platform terms. Ultimately, the paper contributes an audited negative result that proves current LLM evaluation methods are often unrepeatable due to infrastructure-level non-determinism, calling for a more disciplined approach to measurement science in machine learning.
Paper 5
The article evaluates 'Random Attention' as an efficient KV cache eviction strategy for large language model reasoning tasks.
TL;DR
This technical report examines the efficacy of using Random Attention for KV cache eviction to optimize LLM inference throughput. The study shows that this method significantly boosts tokens per second in high-concurrency scenarios without sacrificing model accuracy.
The research investigates 'Random Attention' as a lightweight alternative for KV cache eviction, integrated into the TriAttention compression framework. Using an NVIDIA H200 GPU and vLLM with PagedAttention, the authors benchmarked various Qwen3 and Phi-4 models under heavy load. The experiments focused on how different eviction strategies affect throughput and accuracy during long-context generation (up to 32k tokens). Results indicate that Random Attention serves as an effective scoring function, allowing for higher resident sequence counts before hitting preemption limits. Specifically, on Qwen3-4B, the method increased steady-state throughput from 2046 to 2188 tok/s at certain load points and achieved up to a 41% increase in throughput compared to standard methods when operating at the preemption-safe ceiling. The study also notes that while larger models like Qwen3-32B consume more KV pool capacity, reducing the available sequence cap, the Random Attention approach remains robust across different model scales and maintains high accuracy (0.864 on MATH500). The findings suggest that simple random selection for eviction can provide substantial performance gains in memory-constrained inference environments.
Top stories curated from across the web this week
Article 1
The research explores how 'abliterated' (uncensored) AI models exhibit 'verdict bias,' leading to significantly higher false positive rates in automated vulnerability research.
TL;DR
This article investigates the unintended side effects of using abliterated open-weight models for automated bug hunting. It identifies a phenomenon called 'verdict bias,' where removing model refusals causes the AI to confirm vulnerabilities even when its own reasoning proves they are protected.
The author conducts an empirical study on the utility of 'abliterated'—or uncensored—large language models in the context of automated vulnerability research. Using a pipeline designed to scan FreeBSD kernel source code for potential zero-day vulnerabilities, the researcher observed that while abliterated versions of the Gemma model family produced significantly more bug candidates, they were fundamentally less accurate. Specifically, these models demonstrated 'verdict bias,' a tendency to return a positive vulnerability verdict even when the underlying Chain-of-Thought reasoning correctly identifies the presence of security bounds and guards. The study highlights a specific instance where an abliterated model identified a heap overflow in a function, explicitly noted that the buffer was bounded by MAX_AUTH_BYTES, yet still concluded that an overflow would occur. This suggests that the weight-editing techniques used to strip refusals from models may inadvertently degrade the model's ability to maintain logical consistency during decision-making, ultimately increasing the false positive rate and the manual triage burden for security researchers.
Article 2
A critical analysis of a recent peer-reviewed paper by D-Wave and Japan Tobacco regarding quantum-assisted molecular design.
TL;DR
The article critiques a D-Wave and Japan Tobacco paper, arguing that the 'quantum' contribution to molecular design is minimal compared to classical algorithmic improvements. It highlights significant methodological flaws and positions D-Wave as being behind competitors in practical quantum chemistry applications.
The Coherence Report provides a technical critique of a recent publication in Scientific Reports by D-Wave and Japan Tobacco concerning quantum-assisted molecular design. While the paper introduces a legitimate classical contribution via a new 'Neural Hash Function' for training models with discrete representations, the author argues that the quantum component—using the D-Wave Advantage2 processor as a sampler—offers negligible benefits. The quantum annealer is used solely to generate fingerprints in an autoencoder architecture, meaning no actual chemical properties or electronic structures are encoded on the QPU. The analysis points out several scientific weaknesses: the reported improvements are inconsistent across different architectures (Transformer vs. MLP), the study lacks variance reporting due to single training runs, and the benchmarks (Validity and Drug-likeness) are superficial syntax checks that ignore critical metrics like novelty and synthetic accessibility. Furthermore, the report notes a lack of external benchmarking against modern industry standards. Finally, the article contextualizes D-Wave's position within the broader quantum landscape, noting that competitors such as IBM, Quantinuum, and Pasqal have established much deeper integrations with major pharmaceutical entities like Moderna, AstraZeneca, and Sanofi through more chemically significant hardware applications.
Article 3
Google introduces TimesFM-3, a new 330M parameter foundation model designed for zero-shot multivariate time-series forecasting.
TL;DR
Google has announced TimesFM-3, an evolution of its time-series foundation model that moves beyond univariate forecasting to support multivariate scenarios. The model leverages 330 million parameters and a trillion-point dataset to predict multiple coevolving series using both past and future covariates.
The announcement marks a significant milestone in the evolution of time-series foundation models, moving from the univariate limitations of previous iterations like TimesFM-2.5 to the multivariate capabilities of TimesFM-3. While earlier models focused on forecasting based solely on the history of a single series, TimesFM-3 is designed to handle the complexity of real-world scenarios where multiple variables interact. The model architecture utilizes 330 million parameters and has been trained on an extensive dataset comprising over one trillion time points from both synthetic and real-world sources. A key feature of this new generation is its ability to perform zero-shot forecasting across complex dependencies, meaning it can provide accurate predictions for new tasks without requiring retraining or fine-tuning. Specifically, the model supports multiple targets, allowing for the simultaneous forecasting of related time series such as different product categories in retail. Furthermore, it can integrate past covariates, such as historical foot traffic, and dynamic future covariates, such as upcoming weather forecasts or planned promotional events. This capability allows the model to capture the intricate relationships between coevolving variables, significantly improving forecast accuracy in domains ranging from finance and manufacturing to healthcare and natural sciences.
Article 4
A comprehensive review of experimental claims regarding quantum computational advantage and the ongoing competition between quantum and classical computing capabilities.
TL;DR
This review article provides a detailed survey of the landscape of quantum computational advantage, examining both successful experimental claims and the classical algorithmic advancements that challenge them. It emphasizes the transition from noisy intermediate-scale quantum (NISQ) utility to the necessity of error correction for true algorithmic supremacy.
The article presents a rigorous historical and technical overview of the race between quantum and classical computation. It begins by cataloging significant experimental milestones where quantum processors, such as superconducting and photonic systems, have claimed computational advantage through tasks like random circuit sampling and Gaussian boson sampling. However, the review maintains a critical perspective by documenting how advancements in classical computing—including the use of massive GPU clusters and specialized supercomputers like Sunway—have frequently closed the gap, effectively redefining the baseline for what constitutes 'quantum supremacy.' Beyond hardware-specific benchmarks, the text explores theoretical computational advantages in practical domains such as approximate optimization and recommendation systems. A significant portion of the discussion is dedicated to the current limitations of NISQ-era devices and the critical importance of quantum error correction. The author posits that while 'utility' can be demonstrated in the current noisy era, reaching the transformative potential of algorithms like Shor's algorithm depends entirely on overcoming the frontier of error-corrected quantum computing.
Article 5
Researchers at PolyU have developed a 2D quantum-tunnelling field-effect transistor (TFET) that overcomes the physical energy efficiency limits of conventional MOSFET technology.
TL;DR
A research team led by Hong Kong Polytechnic University has engineered a novel 2D heterostructure transistor that bypasses the fundamental energy efficiency limits of traditional semiconductors. This breakthrough promises much lower power consumption for next-generation AI chips and high-performance computing.
Modern semiconductor progress is currently stalled by the 'Boltzmann tyranny,' a physical limit in conventional MOSFET technology that restricts subthreshold swing to a minimum of 60 millivolts per decade at room temperature. This limitation prevents further reductions in power consumption for high-performance electronics. To address this, a collaborative research team led by Professor Jianhua Hao at The Hong Kong Polytechnic University has developed a Tunnelling Field-Effect Transistor (TFET) using two-dimensional nanomaterials. By utilizing an ultra-thin heterostructure of bismuth (Bi) and indium selenide (InSe) produced through pulsed laser deposition, the researchers successfully replaced thermionic emission with quantum tunnelling. This mechanism allows the device to operate below the 60 mV/decade threshold across six orders of magnitude of current switching. The resulting transistor requires a gate-voltage range of only 160 mV, compared to the 800 mV required by advanced MOSFETs. Furthermore, the device achieves high output current and an exceptional ON/OFF ratio, which are critical for driving downstream logic gates in complex circuits. Because the production method is compatible with existing silicon-based manufacturing processes on centimeter-scale substrates, this technology provides a scalable roadmap for the development of ultra-low-power, energy-efficient microchips essential for the future of artificial intelligence and advanced computing hardware.
Article 6
A new ClickFix campaign is using Polygon blockchain technology via 'EtherHiding' to dynamically update command-and-control servers and evade detection.
TL;DR
A sophisticated ClickFix campaign is leveraging the Polygon blockchain to implement 'EtherHiding,' allowing attackers to rotate C2 servers via smart contracts. This technique makes traditional IP and domain blocking ineffective by using a distributed ledger as an attacker-controlled command center.
A recent investigation by GuidePoint Security's Research and Intelligence Team (GRIT) has uncovered a highly resilient ClickFix campaign that has impacted at least 31 organizations. The defining characteristic of this attack is the use of 'EtherHiding,' a technique where attackers utilize the Polygon blockchain to host and update command-and-control (C2) server addresses. By using smart contracts on a permanent, distributed ledger, the attackers can redirect infected machines to new C2 infrastructure for a negligible cost, rendering standard domain and IP blocking strategies largely obsolete.
The attack vector is two-pronged: first, legitimate business websites (often WordPress-based) are compromised via vulnerabilities or malicious JavaScript injection. Second, end-users are targeted through Search Engine Poisoning. When a user visits a compromised site, the attackers use a malicious JavaScript overlay that mimics CloudFlare's human verification process. This overlay presents the 'ClickFix' lure, instructing users to execute specific keyboard commands (Windows+R, Ctrl+V, Enter) to run a malicious dropper.
This campaign deviates from standard ClickFix infostealer operations by deploying a persistent agent and a dropper that establishes long-term access. The malware is designed to survive reboots and beacons to the C2 every minute, retrieving updated instructions directly from the Polygon blockchain. Researchers suggest that the complexity of this infrastructure points toward an Initial Access Broker (IAB) rather than a typical infostealer operator. To mitigate these threats, security professionals are advised to implement advanced social engineering training and technical controls, such as blocking unauthorized queries to blockchain RPC endpoints and enhancing PowerShell logging and alerting.
Article 7
An analysis of Server-Side Request Forgery (SSRF) vulnerabilities caused by the loss of security bindings during various handoffs between URL validation and socket connection.
TL;DR
The article explores how SSRF vulnerabilities arise when the security verdict of a URL validator is not preserved through the entire request lifecycle. It identifies specific 'handoff' points where discrepancies between parsers and clients can lead to unauthorized access to internal resources.
The author presents a deep technical investigation into the breakdown of security bindings during the process of validating and executing HTTP requests. Using Mealie as a primary case study, the article demonstrates how a validator using an IPv4-only resolver (gethostbyname) can be bypassed when the underlying client (libcurl) uses a dual-stack resolver (getaddrinfo) to connect to an unauthorized IPv6 address. The research extends beyond simple parsing errors to examine six critical handoff stages: URL parsing, object construction, resolution/selection, connection, authority emission, and response handling. Through a large-scale differential analysis of sixteen validators against five HTTP clients across eight programming languages, the author identifies various 'failure families' (F1 through F12). These failures include discrepancies in authority boundaries, differences in how IP addresses are classified, and instances where application state or protocol features like HTTP/2 :authority headers override the validated URL. The article concludes that even when following OWASP mitigations, security is compromised if the application fails to bind the approved destination directly to the socket being opened, particularly in environments utilizing dual-stack DNS, NAT64, or complex redirect logic.
Article 8
The paper introduces Direct Kolen-Pollack Predictive Coding (DKP-PC), a new algorithm designed to accelerate neural network training by reducing error propagation delay.
TL;DR
The researchers present DKP-PC, an optimized version of predictive coding that utilizes direct feedback pathways to eliminate depth-dependent delays. This method significantly improves the efficiency and scalability of biologically inspired neural network training.
Predictive Coding (PC) is a prominent biologically inspired algorithm used for training neural networks through local updates, which facilitates parallel learning across different layers. Despite its advantages, standard PC implementations suffer from two significant bottlenecks: error signals must propagate sequentially from the output layer back to early layers through multiple inference steps, and these signals experience exponential decay, leading to vanishing updates in deeper layers. To resolve these issues, the authors propose Direct Kolen-Pollack Predictive Coding (DKP-PC). This new approach integrates direct feedback alignment and the direct Kolen-Pollack algorithm to create learnable connections that link the output layer directly to all hidden layers. By establishing this direct pathway for error transmission, the complexity of error propagation is reduced from O(L) to O(1), effectively removing the delay associated with network depth. Empirical evaluations show that DKP-PC maintains or exceeds the performance of standard PC while providing superior latency and computational efficiency. These improvements make DKP-PC a highly promising candidate for implementation on custom, hardware-efficient architectures.
Article 9
A new procedure for transforming stabilizer codes into Floquet codes using distance-preserving ZX-calculus rewrites.
TL;DR
Researchers have developed a method to transform complex stabilizer codes into more efficient Floquet codes using ZX-calculus. This process utilizes distance-preserving rewrites to ensure that the new codes maintain their error-correcting capabilities while simplifying hardware implementation.
The paper addresses the practical challenges of implementing stabilizer codes with large weight measurements in fault-tolerant quantum computing. To mitigate the difficulty of performing high-weight operations, the authors propose a 'Floquetification' procedure. This method synthesizes novel Floquet codes from existing stabilizer codes by decomposing complex measurements into sequences of single- and two-qubit operations. A significant technical hurdle addressed is that standard ZX-calculus rewrites often fail to preserve code distance; therefore, the authors define and implement 'distance-preserving rewrites' to guarantee that the resulting Floquet code retains the original code's distance and number of logical qubits. While this transformation introduces a linear overhead in physical qubits based on the weight of the largest measurement, it allows for more efficient data encoding and easier computation. The methodology is grounded in the ZX-calculus graphical language and provides a scalable framework for generalizing Floquet code construction to arbitrary stabilizer codes, with potential implications for broader fault-tolerant quantum computation design.
Article 10
Research indicates that quantum computers require significantly fewer resources to break blockchain-specific elliptic curve cryptography compared to RSA systems, threatening the security of cryptocurrencies and critical infrastructure.
TL;DR
New research reveals that the smaller key sizes used in blockchain elliptic curve cryptography make them uniquely vulnerable to quantum computing attacks. This threat extends from individual cryptocurrency wallets to systemic risks involving cross-chain bridges and real-world asset tokenization.
The article explores the growing disparity between the security of traditional RSA systems and the elliptic curve cryptography (ECC) used in blockchain networks when faced with quantum computing advancements. Because ECC keys are significantly smaller than RSA keys for comparable security levels, they require fewer quantum resources to compromise via Shor's algorithm. This vulnerability is particularly acute in Bitcoin's P2PK scripts, where public keys are explicitly exposed on the ledger, and in protocols like Mimblewimble that rely on ECDH and Pedersen commitments. The research highlights that simply increasing key sizes may only offer temporary protection as quantum algorithms evolve. Furthermore, the shift toward post-quantum signature schemes like Falcon introduces new complexities, including potential side-channel attacks and increased computational overhead for decentralized networks. Beyond simple theft of funds, the article warns of systemic risks to the broader DeFi ecosystem, including the potential for attackers to manipulate cross-chain bridges, collapse stablecoin pegs, and undermine the integrity of tokenized real-world assets (RWAs). The findings emphasize an urgent need for well-vetted cryptographic transitions and robust key rotation mechanisms to mitigate the looming threat of cryptographically relevant quantum computers (CRQCs).
Article 11
A human attacker utilized AI agents and agentic frameworks to execute a highly efficient ransomware attack that breached an enterprise network in under 10 hours.
TL;DR
An enterprise network was breached in record time by a human attacker using autonomous AI agents to automate the entire attack chain. The intrusion concluded with the attackers leaving an extensive 80-page security audit detailing the exploited vulnerabilities.
A recent investigation by Unit 42 revealed a groundbreaking ransomware attack characterized by extreme operational efficiency through the use of frontier AI models and agentic frameworks. A human operator orchestrated a series of automated tasks where AI agents performed reconnaissance, breached a public API endpoint, and tunneled into the enterprise network. Once inside, specialized subagents were deployed to map internal microservices and scrape code repositories for hard-coded tokens and service passwords. By leveraging these stolen credentials, the attackers accessed secret-management systems to obtain master administrative credentials, granting them root access across the organization's cloud, identity, CI/CD, and SaaS environments. A particularly sophisticated move involved hijacking CI/CD workflows to convert the victim's own cloud AI services into infrastructure for the attacker, allowing them to hide orchestration traffic within legitimate compute activity. In a highly unusual post-attack behavior, an AI agent left the victim with an 80-page security audit documenting dozens of exploited findings. To combat such machine-speed threats, security experts recommend that organizations deploy automated defensive playbooks capable of revoking credentials and isolating accounts in real-time. Furthermore, companies must treat AI as critical infrastructure by auditing all model endpoints, API keys, and MCP gateways while implementing strict rate limits and least-privilege policies to prevent both unauthorized access and massive token consumption costs.
Article 12
Anthropic researchers present new mechanistic interpretability techniques to observe and understand the internal computational processes of large language models like Claude.
TL;DR
Anthropic has introduced new research methods to map the internal 'circuits' of Claude, revealing how it processes information. The study provides empirical evidence for advanced behaviors like cross-lingual conceptual sharing and predictive planning in text generation.
In a significant advancement for mechanistic interpretability, Anthropic researchers have unveiled two papers detailing progress in observing the internal mechanics of large language models (LLMs). By treating the model's weights and activations as a biological system to be studied via an 'AI microscope,' the team has moved beyond analyzing mere outputs to mapping the underlying computational circuits. One primary finding is the existence of a shared conceptual space across multiple languages, implying that Claude processes abstract meanings in a language-agnostic manner before translating them into specific tongues. Furthermore, the research debunks the idea that models only predict the next token without foresight; specifically, in poetic tasks, the model demonstrates 'planning ahead' by selecting rhyming words before completing the sentence structure. The researchers also identified risks associated with model transparency, such as the ability to detect instances where the model fabricates reasoning to please a user or follows incorrect hints. While current methods are computationally expensive and only capture a fraction of total model activity, these tools represent a critical step toward ensuring AI alignment, safety, and reliability by making the 'black box' of neural networks more transparent.
Article 13
An experimental demonstration of using Claude Code (Claude Opus) to reverse engineer a zero-day vulnerability in Papercut NG and develop a functional exploit chain.
TL;DR
This article details a highly technical experiment where an LLM-driven agent was used to automate the discovery and exploitation of vulnerabilities in Papercut NG. By leveraging patch diffing and automated infrastructure, the AI developed a full proof-of-concept for unauthenticated code execution.
The author describes a controlled experiment using Claude Code to test 'N-Day' research capabilities against a recently patched vulnerability in Papercut NG. Utilizing a custom research harness equipped with tools like Ghidra, semgrep, and Playwright, the AI agent was tasked with replicating the vulnerability and building an exploit chain. Within the first few minutes, the agent identified SQL injection indicators in the logs and used patch diffing to pinpoint changes in the Java bytecode related to Apache Derby system stored procedures. The research progressed to identifying an authentication bypass caused by a mismatch in path-normalization between security filters and the Spring dispatcher, specifically using matrix parameters to evade checks. Despite encountering Anthropic's safety guardrails during the verification phase, the agent successfully transitioned models and continued the process, ultimately achieving a functional proof-of-concept for unauthenticated code execution within approximately 90 minutes of active processing.
Article 14
The article explores 39 documented attack vectors that bypass the cryptographic security of passkeys by targeting the surrounding ecosystem, such as operating systems, cloud synchronization, and enrollment processes.
TL;DR
While passkeys offer strong cryptographic protection against phishing, researchers have identified 39 methods to bypass this security by targeting the surrounding ecosystem. These attacks focus on exploiting operating systems, synchronization services, and enrollment workflows rather than breaking the underlying protocol.
The article presents a critical analysis of the evolving threat landscape surrounding passkey authentication. Although FIDO2-based passkeys are designed to be phishing-resistant through public key cryptography, the author highlights that 39 documented attack paths exist that target the broader ecosystem rather than the cryptographic core. These vulnerabilities include 'assertion replay' and 'browser hooking,' where malware leverages legitimate WebAuthn infrastructure to sign unauthorized requests. Furthermore, the convenience of synchronized passkeys—which move between devices via cloud services like Apple or Google—introduces risks from compromised password managers and mobile OS vulnerabilities. The research also points to significant weaknesses in enrollment and recovery, such as 'shadow passkey' creation and help desk takeover, where attackers register their own credentials rather than stealing existing ones. To counter these architectural flaws, the article suggests that enterprises should consider dedicated biometric hardware. Such devices minimize the attack surface by removing general-purpose operating systems, browsers, and synchronization capabilities, thereby making it much harder for attackers to manipulate the user interface or intercept authentication assertions.
Article 15
The Vortex columnar file format aims to optimize machine learning training by accelerating data loading from S3 directly to GPUs.
TL;DR
This article explores how the Vortex columnar file format optimizes the data pipeline for ML training by reducing latency between cloud storage and GPUs. It highlights techniques like column pruning and GPU-side decompression to eliminate traditional CPU and NVMe bottlenecks.
The presentation by Onur Satici of SpiralDB addresses the critical inefficiency in modern machine learning training pipelines, specifically the 'movement tax' incurred when transferring data from cloud storage like S3 to GPUs. Current workflows often rely on downloading data to local NVMe storage, decompressing it via the CPU, and then transferring it over PCIe, which creates significant bottlenecks in both throughput and latency. To solve this, the Vortex file format introduces a highly optimized columnar structure designed for high-speed streaming. Unlike Parquet, Vortex decouples logical types from physical encodings and utilizes lightweight, cascading encodings that permit operations like aggregation and filtering to be performed directly on compressed data. This approach allows for massive throughput—demonstrated by a 4K video visualization processing at 13 gigabits per second—and significantly faster random access compared to Parquet. The format's architecture relies on a hierarchical layout system, including ZonedLayouts that use zone-maps (min/max statistics) to enable efficient segment pruning. By incorporating alignment information directly into the file segments, Vortex enables zero-copy transfers and optimized SIMD or GPU kernel execution, effectively transforming the data loading process from a multi-step bottleneck into a streamlined, high-bandwidth pipeline.
Article 16
The article discusses the development and advantages of quantum neural operators, specifically focusing on their implicit quadratic frame and expressivity benefits in solving complex differential equations.
TL;DR
This technical compilation explores the intersection of quantum computing and neural operators for solving partial differential equations. It highlights how quantum architectures provide superior expressivity and computational advantages through implicit quadratic frames.
The provided text is a comprehensive bibliography and reference list centered on the advancement of Quantum Neural Operators (QNOs). The core research focus is on utilizing quantum computing to enhance the efficiency and expressive power of neural operators, such as DeepONet and Fourier Neural Operators, when addressing nonlinear partial differential equations. Key technical themes include 'Schrödingerization' for simulating PDEs on quantum hardware, the use of differentiable quantum circuits for solving nonlinear equations, and the development of physics-informed quantum machine learning models. The references trace the evolution from classical operator learning to recent breakthroughs in quantum-accelerated frameworks like Quantum DeepONet and QuanONet. Furthermore, the text touches upon the practical implementation challenges within the NISQ era and mentions specific software frameworks like MindSpore Quantum and DeepXDE that facilitate these complex simulations. Ultimately, the research aims to demonstrate that quantum neural operators possess fundamental mathematical advantages in terms of their approximation capabilities and computational complexity when compared to their classical counterparts.
Article 17
The European Commission's designation of OpenAI's ChatGPT as a Very Large Online Search Engine (VLOSE) under the Digital Services Act.
TL;DR
The European Commission has officially designated ChatGPT as a Very Large Online Search Engine (VLOSE) under the Digital Services Act. This move subjects OpenAI to rigorous transparency, auditing, and risk-mitigation requirements regarding its search and advertising functions.
The European Commission's decision to designate ChatGPT as a Very Large Online Search Engine (VLOSE) marks a significant regulatory milestone for generative AI in the EU. By classifying the service under this stringent tier of the Digital Services Act (DSA), the Commission has demonstrated the law's technological neutrality, applying existing search engine frameworks to new, hybrid AI interfaces. This designation means OpenAI must now undertake comprehensive risk assessments and implement mitigation measures to address systemic threats, such as impacts on democratic processes, public health, and media pluralism. A critical component of this regulation is the increased transparency provided to researchers; under Article 40, vetted researchers can access data to investigate how ChatGPT's algorithmic curation and citation behaviors affect information integrity. Furthermore, the timing of this designation coincides with OpenAI's expansion into advertising within Europe, necessitating the creation of an ad repository. Looking forward, while these specific obligations do not take full effect until January 2027, the VLOSE status sets a precedent that could eventually lead to ChatGPT being designated as a 'gatekeeper' under the Digital Markets Act (DMA) if it meets the necessary quantitative thresholds.
Article 18
AI coding agents are inadvertently executing untrusted code from unverified sources, creating new supply chain vulnerabilities in corporate networks.
TL;DR
Security researchers have demonstrated that AI coding agents are executing unverified code found in documentation files on corporate networks. This behavior exposes major enterprises to sophisticated supply chain attacks by exploiting the inherent trust agents place in vendor-provided instructions.
A recent investigation by a security startup has revealed a critical flaw in how AI coding agents interact with external documentation. By scanning over 6,000 domains belonging to high-value targets like defense contractors and Fortune 500 companies, researchers found numerous instances where 'llms.txt' files pointed to unregistered or unclaimed code packages and domains. To validate the risk, researchers registered these unclaimed names and hosted malicious beacons designed to trigger a response when executed. The experiment was successful, resulting in 'phone-home' signals from multiple large corporations within an hour. The investigation traced the execution chain back to popular AI agents, including Anthropic's Claude, OpenAI's Codex, and Nous Research's Hermes. The core of the issue lies in a broken trust model: these autonomous agents are programmed to treat documentation as absolute truth and do not perform the necessary verification of the underlying code or domains referenced within those documents. As agentic AI usage expands across SaaS, cloud, and endpoint layers, this creates an expanding supply-chain attack surface that current security guards are not equipped to monitor or defend against.
Article 19
QuFi Network has launched a post-quantum verification platform designed to provide quantum-resistant validation for digital assets without altering existing settlement networks.
TL;DR
QuFi Network has launched a new post-quantum verification layer to secure digital assets against future quantum computing threats. The platform utilizes advanced NIST-standardized cryptographic algorithms to provide lightweight proofs that can be validated without modifying underlying blockchain settlement layers.
As the transition toward post-quantum cryptography (PQC) becomes an economic imperative, QuFi Network has introduced a specialized verification infrastructure to address the scalability challenges of quantum-resistant signatures. Traditional blockchain networks face significant overhead because PQC signatures can be over 100 times larger than current elliptic-curve standards, leading to increased storage and bandwidth costs. The QuFi Platform solves this by acting as a neutral, decoupled layer that performs cryptographic validation upstream from the settlement process. By separating verification from settlement, the platform allows for complex computations to occur in a dedicated environment, delivering compact proofs to downstream systems like Bitcoin or EVM-compatible chains.
The technical architecture of the platform relies on a hybrid approach using FIPS-standardized algorithms: ML-DSA-65 for digital signatures, SLH-DSA as a hash-based backup, and ML-KEM-1024 for secure key encapsulation. This multi-family cryptographic strategy ensures long-term resilience against various quantum attack vectors. To demonstrate the practical utility of this architecture, QuFi has deployed uBTC on Bitcoin Testnet4, which serves as a proof-of-concept for verifying Bitcoin collateral through hybrid signature challenges while maintaining standard Bitcoin transaction finality.
Looking forward, QuFi is expanding its ecosystem through the Genesis Node Program, inviting infrastructure providers, institutional custodians, and enterprise partners to operate independent verification nodes. The roadmap includes integrations with major Layer 1 networks such as Solana and Stellar, aiming to create a shared, scalable infrastructure for the multi-trillion-dollar tokenized asset market.
Article 20
A summary of major US technology policy developments in August 2026, including the Meta child safety settlement, White House AI risk frameworks, and federal data privacy concerns.
TL;DR
The August 2026 tech policy landscape was dominated by Meta's massive settlement regarding teen safety and a secretive White House framework for auditing high-risk AI models. Additionally, significant legal battles emerged over federal agencies bypassing warrants to access private citizen data through commercial brokers and inter-agency sharing.
The month of August 2026 saw landmark shifts in US technology regulation and oversight. A major highlight was the settlement between Meta and 52 state attorneys general, where Meta agreed to pay up to $17.1 billion and implement strict usage controls for teen users to mitigate alleged social media harms. This settlement is viewed as a bellwether for ongoing litigation against social media platforms. In the realm of artificial intelligence, the White House finalized a non-public voluntary framework designed to review the cyber capabilities of closed-source frontier models before release. However, this framework has sparked intense debate due to its exclusion of open-weight models and concerns regarding transparency in government oversight. Simultaneously, significant privacy concerns were raised as federal agencies like the SEC, DHS, and ICE were found to be accessing sensitive personal data—ranging from airline passenger records to taxpayer information—through methods that circumvent traditional warrant requirements. These actions have prompted legal challenges from various state attorneys general seeking to protect citizens' privacy rights against invasive government surveillance and data-sharing practices.
Free weekly digest
The week’s most relevant AI, security, blockchain, and engineering stories — curated, summarised, and reviewed by humans. No spam, unsubscribe anytime.
Subscribe — it’s free