A fundamental flaw leaves LLMs strikingly vulnerable to attack
Researchers claim LLMs have a fundamental, unfixable flaw making them impossible to fully secure against attacks, presented at ICML.
MIT Technology Review: AI·30 Jul
Search signals, briefings, benchmarks and glossary terms.
Search signals, briefings, benchmarks and glossary terms.
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Researchers claim LLMs have a fundamental, unfixable flaw making them impossible to fully secure against attacks, presented at ICML.
LinkedIn is not expanding its data centers in the next year, prioritizing compute efficiency over increased GPU spending for AI initiatives.
OpenAI has reduced pricing for its GPT-5.6 models, Luna and Terra, claiming improved efficiency for enterprise AI workflow deployment.
The EU launched a call for tenders to establish up to seven AI Gigafactories across Europe, aiming to boost computing capacity with over €30 billion in investment.
Researchers found an AI agent (Claude) was more effective than a human at building “exploitable trust” in text conversations over one week.
Chinese stocks linked to AI suppliers are experiencing their worst monthly performance in a decade, as investors withdraw from the sector.
Microsoft reported strong Q4 earnings driven by cloud and AI, while Meta's disappointing revenue forecast and low free cash flow signals rising AI spend.
Australia's eSafety regulator has initiated civil penalty proceedings against Telegram for allegedly failing to remove pro-terror material from its messaging service.
Financial Times examines how in-house legal teams at law firms are creatively adopting AI tools to meet evolving business demands.
The Financial Times reports that investment concentration risk in AI is extending beyond equities into bond markets, driven by a shared investment thesis.
ByteDance is significantly increasing its investment in AI development, with some market observers viewing this strategy as high-risk.
New research proposes a probabilistic K-line forecasting method that ensures consistency by preventing quantile and K-line crossing without retraining.
Researchers propose Think Short, Defer Smart (TSDS), a framework for edge LLM agents to manage reasoning budgets and defer to cloud models based on uncertainty.
Research details an algebraic framework to understand the computational limits of causally masked transformers under finite precision and real-world constraints.
Research explores using graph-based inference to identify business conduct risk when reported incident data is sparse and biased, acknowledging silent risk.
Research explores shared lessons in supervised fine-tuning (SFT) across AI alignment, model organisms, and toy models, testing transferability.
Research introduces 'feature instability' (FI) to measure the stability of feature bagging, an ensemble method for machine learning.
Research investigates context sampling choices for TabPFN on small tabular datasets, impacting prediction stability, accuracy, and selection cost.
Research proposes an uncertainty-guided LLM semantic augmentation method to improve heterogeneous treatment effect estimation in personalized interventions.
Research reveals LLMs generate correlated fictional names, like 'Elena Vasquez' and 'Marcus Chen,' across AI-generated documents.
Research investigates how statistical similarity in train-test splitting impacts AutoML model evaluation, especially with imbalanced datasets.
Researchers introduced a novel statistical framework for evaluating bias in medical imaging machine learning models, using counterfactual analysis to assess dependency on sensitive attributes.
TLA-Prover, a 20B-parameter model, significantly improves verifiable TLA+ specification synthesis, achieving higher semantic model-check rates than other LLMs.
REAP introduces an automatic method for curating coding agent benchmarks from interactive production usage, aiming to improve evaluation speed and fidelity.
ToxScreen, a new research technique, detects backdoors from poisoned training data in LLMs, even without original data or trusted models.
New research proposes Dense Hierarchical Rewards and Curriculum Learning (DHRCL) to improve code LLM training using execution, unit test, and structural analysis.
New research proposes a game-theoretic approach to fine-tuning large language models, balancing performance improvement with drift control via KL regularization.
New research explores 'any-order inference' for AI models, proposing masked diffusion models as a native approach for non-causal reasoning.
Researchers propose FedDAB, a two-phase defense method using local contrastive regularization and alignment checking to counter backdoor attacks in Federated Learning (FL).
Research introduces a method for high-order Markov blanket discovery, relaxing the faithfulness assumption common in causal and graphical model learning.
Research paper introduces Top-k Pareto Bandits, a multi-objective bandit algorithm for selecting a slate of k actions that approximate a Pareto frontier.
New research explores a continuum of LoRA initialization methods, improving fine-tuning performance by refining adapter initialization based on loss gradients.
FloDR, a new invertible dimensionality reduction method based on a normalising flow, addresses limitations of t-SNE and UMAP in preserving data structure.
AgentGFM introduces a Graph Foundation Model with node-agent information-flow control, addressing variability in local graph structural patterns.
Research explores stable and budget-feasible coalition formation for clustered federated learning, using a hedonic potential-game approach.
Researchers propose Weak-to-Strong On-Policy Distillation (OPD) to transfer LLM capabilities, addressing limitations of existing distillation methods.
RAGuard proposes a layered defense framework using adversarial fine-tuning for dense retrievers to protect RAG systems from corpus data poisoning attacks.
New research proposes an Existence-Field Diffusion Model to generate spatial point processes with variable cardinality, overcoming limitations of existing methods.
Deep reinforcement learning agents using frozen random CNNs develop extremely sparse, task-relevant representations without explicit sparsity objectives.
Research introduces conformal changepoint localization for robust root cause analysis in engineered systems, providing statistical reliability guarantees.
Researchers introduced HoF-Bench, a new benchmark of 95 real AI-discovered CVEs, to evaluate LLM-based security analyzers like AISLE's.
Research addresses compounding error in video diffusion models for long-horizon generation, proposing a regularization technique to improve frame quality.
Researchers propose HiFloat4 for the first end-to-end 4-bit precision Reinforcement Learning (RL) post-training of large language models.
Research identifies a structural issue where permutation-equivariant GNN explainers produce non-canonical top-k explanations due to arbitrary ordering.
New research proposes Implicit Causal World Models to learn environmental dynamics and agent intents from multi-agent demonstrations, addressing distribution shifts.
Research presents an analytical method to estimate energy consumption of LLM inference on GPUs, aiming to simplify sustainability reporting and design.
New research proposes a nongradient vector flow method for learning flow maps in diffusion and consistency models, aiming to reduce inference overhead.
New research proposes a method for online conformal prediction that simultaneously guarantees coverage and efficiency, addressing limitations of existing adaptive conformal inference techniques.
New research proposes Sequential Adaptive Rollout Allocation (SARA) to improve compute efficiency in Reinforcement Learning with Verifiable Rewards (RLVR).
Research paper clarifies that dynamic parameterization in AI models does not equate to dynamic inference or computational savings, emphasizing a 'frozen-controller auditing' principle.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion