Search: “tee”

"They don't care about this": A Systematic Study of TEE Build Reproducibility in the Wild

91% of 115 surveyed TEE deployments across Intel SGX, TDX, and AMD SEV fail to provide reproducible builds needed for verifiable remote attestation.

A systematic study of 115 TEE deployments found 91% were not reproducible and 80% lacked both source code and a reference build, undermining remote attestation guarantees. Interviews with 12 developers of 50 Intel SGX projects confirmed that only one participant treats reproducibility as a development priority. The authors identify technical barriers such as embedded timestamps plus ecosystem-level issues like lack of build-environment control in multi-stakeholder projects, and call for holistic, committed reproducibility practices.

arXiv cs.CR·Sep 10, 11:43 UTC · 6d agoResearch

TCRF taken offline by DDoS attack after Claude user ban

The Cutting Room Floor game wiki was taken offline by a DDoS attack after a user leveraging Claude was banned.

The Cutting Room Floor (TCRF), a wiki documenting unused video game content, was knocked offline by a distributed denial-of-service attack. The attack reportedly followed moderation action banning a user who was using Anthropic's Claude. The incident highlights friction between community sites and AI-assisted users and tools.

Lobsters · security·Aug 28, 21:26 UTC · 19d agoAI safety & security in the wild

Hoss: Fast Oblivious Semantic Search with Heterogeneous GPU-CPU-TEE Architecture

Hoss uses heterogeneous GPU-CPU TEEs for oblivious semantic search, achieving up to 67x speedup over the Compass ORAM-based system.

Hoss is a first-of-its-kind oblivious semantic search system built on a heterogeneous CPU-GPU TEE architecture, hosting hot-path HNSW graph traversal in large GPU TEE private memory and offloading lower graph layers to CPU TEEs. It invokes oblivious primitives only for lower-layer accesses and adds host-access ORAM and data-dependent optimizations. Benchmarked against prior state-of-the-art Compass, the prototype achieves up to 67x speedup while maintaining high recall, with larger gains at scale.

arXiv cs.CR·Sep 3, 22:21 UTC · 13d agoResearch

You've Got a BUD in Me: Authenticated Reads from Per-Block Write Logs

Researchers propose BUD, per-block write-log digests enabling blockchain validators to serve historical membership and exclusion proofs far cheaper than state-wide tries.

The paper introduces Block Update Digests (BUD), which authenticate each block's write log with predecessor pointers, plus a SuperBUD and exponential hierarchy to turn long unchanged intervals into short proofs. Soundness against adversarial provers and up to f Byzantine validators is proven under archive, attestation, and committee evidence assumptions. Benchmarks show a 50x state-size increase raises the base-BUD path only 1.24x versus 3.1x for in-memory and 69.5x for disk-backed Merkle Patricia tries, with read payloads below 800 bytes and p99 warm verification at 146 microseconds.

arXiv cs.CR·Sep 10, 08:46 UTC · 6d agoResearch

Understanding the Security Boundary of Obfuscation-based On-Device LLM Protection

Researchers formalize obfuscation primitives for TEE-protected on-device LLMs and show a Collapse attack breaks ArrowCloak, TSQP, and LoRO, then extend the boundary.

The paper formalizes obfuscation primitives for TEE-Shielded LLM Partition (TSLP) schemes that offload computationally intensive layers from a Trusted Execution Environment to external GPUs. A novel primitive-guided attack, Collapse, demonstrates a shared vulnerability in prominent published methods including ArrowCloak (Security'25), TSQP (S&P'25), and LoRO (NeurIPS'25). The authors then introduce two new obfuscation primitives and integrate them with existing constructs to formulate an extended security boundary (O_ext).

arXiv cs.CR·Sep 9, 12:51 UTC · 7d agoAI safety & security

FLAT: Resampling Image and Text into 1D Flexible-Length Aligned Transmodal Tokens for Retrieval and Generation

FLAT jointly trains a multimodal encoder with text-to-image and image-to-text decoders, producing flexible-length tokens that hit 83.1 GenEval on T2I after fine-tuning.

FLAT (Flexible-Length Aligned Transmodal representations) is a pre-training framework that jointly optimizes a shared multimodal encoder with T2I and I2T decoders, combining contrastive alignment with bidirectional cross-modal generative objectives. It maps visual and textual inputs into a unified continuous 1D sequence space and uses nested dropout over prefix-K tokens for dynamic output lengths. A single pre-training stage supports cross-modal retrieval and generation (71.1 GenEval), with task-specific fine-tuning reaching 83.1 GenEval on T2I, 40.5 BLEU-4 and 138.6 CIDEr on MS-COCO captioning, and strong Recall@5 on MS-COCO and Flickr30K.

Hugging Face daily papers·Sep 14, 20:00 UTC · 2d agoAI research

SpliTEE: Improving LLM Inference on Trusted Hardware with Differentially Private GPU Outsourcing

SpliTEE splits LLM inference between Intel TDX trusted execution and untrusted GPUs, using differential privacy instead of encryption to protect intermediate representations.

SpliTEE extends split inference to LLMs, running inference partly inside an Intel TDX TEE while masking intermediate inputs sent to untrusted GPUs with differential privacy rather than encryption. The authors show a prompt-reconstruction attack recovers nearly 80% of prompts from unmasked intermediate representations, motivating the masking. A global sensitivity analysis bounds the required DP noise scale, avoiding quantization and keeping models in floating point. The implementation is nearly twice as fast as full CPU-based TDX inference and 5-15 seconds faster than encryption-based Slalom with higher accuracy, evaluated on Llama-3.2-3B and Qwen3-4B.

arXiv cs.CR·Sep 14, 04:55 UTC · 2d agoResearch

ToxicRAG: Compromising Retrieval-Augmented Generation Systems via Single-Shot Knowledge Poisoning Attacks

ToxicRAG shows a single narrative-form poisoned document can steer RAG answers, achieving 0.61-0.91 attack success rates across four LLMs.

The attack injects one document per target question written as a coherent knowledge-update narrative that acknowledges the previously accepted answer, introduces fabricated events that appear to invalidate it, and attributes the attacker-chosen answer to purported authorities. An optional answer-focused self-validation loop revises candidates when a surrogate LLM fails to reproduce the target answer. Across 100 target questions each from Natural Questions, HotpotQA, and MS-MARCO, with four victim LLMs and four dense retrievers, ToxicRAG achieves attack success rates of 0.61-0.91 and matches or exceeds the strongest baseline by 0 to 11 percentage points.

arXiv cs.CR·Sep 10, 04:42 UTC · 6d agoAI safety & security1

Forgery of C2PA on a Pixel 10

Researcher forged a Google Pixel 10 C2PA content credential with genuine signatures, showing root-level attackers can fake photo provenance.

A Hacker Factor blog post demonstrates an AI-generated 'unicorn glitter milk' news photo carrying a valid, cryptographically signed C2PA manifest traceable to Google's Pixel camera certificate chain, passing validation in Adobe Inspect and the CAI Verify tool with a verified timestamp. The author, working with UMBC's PASAWG working group, reported to Google and C2PA in November 2025 that root access on a Pixel device could sign arbitrary images as camera captures; after 90 days without resolution, details were published. The finding undermines C2PA Assurance Level 2 claims made for Pixel 10 Content Credentials.

Lobsters · security·13:24 UTC · 9h agoResearch

Distill Globally, Adapt Locally: Reasoning Distillation and Product-Type Test-Time Training for Scalable Trade-Up Recommendation

A distillation framework compresses LLM reasoning into a 15.5M-parameter trade-up recommendation model reaching AUC 0.941 with product-type test-time training.

The paper targets trade-up recommendation, which identifies higher-quality alternatives that preserve customer purchase intent. A retrieval-augmented few-shot LLM teacher generates labels and rationales that supervise a compact embedding-pair classifier; at inference the 15.5M-parameter student uses only two precomputed 768-dimensional embeddings with no LLM calls. On 8,352 annotated pairs, label-only training scored AUC 0.912, reasoning distillation reached 0.924, and product-type test-time training lifted it to 0.941 with average precision 0.940. The distilled student is roughly 5,000x faster and 10,000x cheaper than direct LLM inference on a 100K-pair proxy catalog.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:08 UTC · 12d agoAI research

Senate set to debate package of bills on privacy, AI and kids safety

CyberScoop·Aug 4, 13:33 UTC · Aug 4, 2026Exploit / PoC in the wild1

Technical Manual for a Toolkit for Measuring Contextual Individuation in Transformer Language Models

An open methodology toolkit measures whether transformer language models contextualize fixed word forms across domains using bridge forms and layer-wise silhouette analysis.

The manual documents an open toolkit built around 'bridge forms' - identical written words recurring across two or more subject domains with a different sense in each - to test whether transformer language models individuate word occurrences by context beyond the embedding layer. It covers declarative specification of bridge forms, Wikipedia corpus acquisition, occurrence localization, layer-wise representation extraction, domain-pairwise silhouette measurement, and visualization, justifying each choice against failure modes such as sense contamination and subword-tokenization misalignment. It is a methodological and implementation reference and reports no empirical results.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 16:38 UTC · 12d agoAI research

DDRop Attack Forces Intel TDX Confidential VMs Into Debug Mode and Exposes Memory

KU Leuven and ETH Zurich researchers released DDRop PoC hardware that forces Intel TDX confidential VMs into debug mode and exposes plaintext memory.

Researchers from KU Leuven, ETH Zurich, Google, and Durham University published proof-of-concept code, hardware designs, and firmware for DDRop, a DDR5 interposer that injects parity errors to drop selected cache-line writebacks. Because Intel TDX, Intel Scalable SGX, and AMD SEV-SNP lack per-line cryptographic freshness, processors decrypt and accept stale DRAM contents as current state. The PoC flips a victim's ATTRIBUTES.DEBUG flag to enable TDX debug mode, then copies victim memory in plaintext and can forge attestation reports. Affected environments include Intel 5th- and 6th-generation Xeon Scalable with TDX; Intel says the attack falls outside its cloud-computing threat model.

GBHackers·Sep 15, 09:57 UTC · 1d agoResearch

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled pure-autoregressive testbed shows task-specific validation losses rank image tokenizers differently, with I2T loss the most consistent signal.

Researchers built a controlled pure-autoregressive testbed and tracked task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They find losses should be analyzed per task because they exhibit distinct scaling behavior and rank tokenizers differently, and that the loss-performance relationship depends on the predicted token space. I2T loss, computed over a shared text vocabulary, correlates consistently with both generation and visual understanding performance after supervised finetuning. Case studies revisit the discriminator, semantic supervision, and vocabulary size as tokenizer design axes.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:57 UTC · 8d agoAI research1

New hardware device can RAM into encrypted memory, expose your data

Researchers built a $200 DDR5 interposer that silently drops memory writes to break TDX, SGX, and SEV-SNP confidential VM integrity, requiring physical access.

Researchers from KU Leuven, ETH Zurich, Durham University, and Google demonstrated DDRop, a hardware interposer costing under $200 that corrupts DDR5 bus commands to silently drop writes to encrypted memory, enabling replay attacks on confidential VMs. Because scalable memory encryption lacks freshness checks, protected VMs keep computing on stale attacker-selected data; on an Intel TDX server the attack forces debug mode for plaintext memory reads or forges attestation reports, succeeding in under two minutes without crashing. Intel and AMD both called the attack out of scope for their cloud threat models, with no mitigation planned, and proposed cache line versioning appears still vulnerable. The full interposer design is being released as open-source hardware.

The Register · Security·Sep 14, 18:31 UTC · 2d agoResearch

Transfer Learning for Evolving Domains

TrED formalizes transfer learning for domains whose data availability evolves over time, arguing classical settings are regimes along one trajectory, and remains unsolved.

The paper introduces Transfer Learning for Evolving Domains (TrED), formalizing transfer learning as a trajectory problem where target data and labels are progressively collected. TrED is specified by a data availability process fixed by the environment, a freely chosen learning protocol, and an evaluation criterion scoring the whole trajectory of models. Classical settings like domain generalization, domain adaptation, and multi-domain learning are recovered as regimes within this framework. The authors survey the literature and find most methods are tailored to a single regime, leaving TrED a well-posed open problem.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 16:37 UTC · 5d agoAI research

Pixel Modem Zero-Day Exploited in Targeted Attacks

Google patched Pixel modem zero-day CVE-2026-58704, a zero-click permission bypass enabling remote privilege escalation, exploited in targeted attacks.

CVE-2026-58704 is a high-severity logic error in the Pixel cellular modem allowing remote (proximal/adjacent) escalation of privilege with no user interaction or additional execution privileges. Google reports limited, targeted exploitation but has not attributed it; the zero-click modem nature suggests commercial spyware or state-sponsored actors. The September Pixel update also fixes more than 100 other Pixel-specific vulnerabilities, nearly 50 of them critical, enabling RCE or privilege escalation.

SecurityWeek·updated 17:56 UTC · 5h ago·first 13:10 UTC · 9h agoExploit / PoC in the wild 8 sourcesCVE-2026-58704

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled autoregressive testbed shows validation losses must be analyzed per task, and image tokenizer choice affects joint multimodal text modeling.

Researchers built a pure-autoregressive testbed to study image tokenizers as the 'visual language' of unified multimodal models, tracking task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They found that losses exhibit distinct scaling behavior per task and rank tokenizers differently, and that I2T loss over a shared text vocabulary gives a more consistent loss–performance signal than T2I loss. Better reconstruction does not necessarily yield lower task-specific losses or stronger downstream performance, and tokenizer choice can affect text modeling under joint optimization. Case studies examine the discriminator, semantic supervision, and vocabulary size design axes.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research1

Session Attestation for Unmodified TLS Services in Confidential Virtual Machines

SessionLatch adds session attestation to unmodified TLS services in confidential VMs, cutting short-upload latency 63.1% versus TNG on Hygon CSV.

SessionLatch provides session attestation for confidential virtual machines without modifying applications, TLS implementations, or certificates. It relies on trusted observation of the server's locally generated ephemeral public key combined with standard TLS key confirmation, moving attestation integration to the operating system via a temporary latch that overlaps evidence exchange with the TLS handshake. Implemented with Linux and Windows integrations and evaluated on real Hygon CSV attestation, it reduces short-upload mean latency by 63.1%/23.0% relative to TNG in interleaved Linux/Windows experiments, while preserving enterprise service authentication and the native TLS data path, and supports mutual attestation.

arXiv cs.CR·Sep 9, 03:36 UTC · 7d agoResearch

TART: A Modular Tool for Technique-Aware Audio-to-Tablature Guitar Transcription

TART pipeline transcribes guitar audio to tablature with fingering and technique labels, outperforming prior baselines by up to 8.5 Tab F1 points.

TART is a modular four-stage pipeline for audio-to-tablature guitar transcription, combining an audio-to-MIDI transcription model, an expressive technique classifier, an audio-conditioned T5 encoder-decoder for string-fret assignment, and an automated tablature generator. In zero-shot evaluation on GuitarSet, EGDB, and two noisy augmented benchmarks, it averages 81.35% audio-to-MIDI F50 (+6.67 over the best baseline), 71.8% string-fret Tab F1 (+8.5), and 54.08% end-to-end Tab F1. The authors state it is the first framework to generate tablature with both fingering and expressive technique annotations directly from guitar audio.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:55 UTC · 6d agoAI research2

Cybersecurity jobs available right now: June 16, 2026

Help Net Security rounds up cybersecurity job openings at Boeing, Lockheed Martin, Siemens Energy, and other firms across the US, Europe, India, and Israel.

Help Net Security's June 16, 2026 roundup lists cybersecurity job openings including Android vulnerability researcher, application security engineer, vulnerability management specialist, DevSecOps engineer, and OT/ICS consultant. Employers include Byteria, Millennium, Nava, Siemens Energy, Lockheed Martin, Enterprise Products, Audubon Companies, Boeing, and General Dynamics across the US, Ireland, Germany, India, Israel, and the UK.

Help Net Security·Aug 17, 09:05 UTC · Aug 17, 2026Industry

Breaking the 1.58-bit Barrier for Ternary LLMs

An arXiv paper claims a method that breaks the 1.58-bit barrier for ternary large language models.

The arXiv preprint 2609.16338, titled 'Breaking the 1.58-bit Barrier for Ternary LLMs,' presents research on ternary-weight large language models, which use roughly 1.58 bits per weight. The source text contained only the title and Hacker News engagement data (56 points, no comments), so further technical details are not available.

Hacker News · AI·20:59 UTC · 2h agoAI researchHN 56↑ · 0 comments

The Machine With Many Faces: Post-Exploitation Identity Misuse in SPIFFE/SPIRE

Unit 42 demonstrates that root access on a Kubernetes node lets attackers spoof SPIFFE/SPIRE attestation and harvest co-located workloads' SVIDs.

Palo Alto Networks Unit 42 describes post-exploitation techniques in which an attacker with root on a compromised Kubernetes node spoofs Linux cgroup metadata used by the SPIRE agent during workload attestation, tricking it into issuing a co-located workload's SPIFFE Verifiable Identity Document to an attacker-controlled process. The research shows the core trust assumption of machine-identity systems—that the node is trusted—collapses once root is obtained, exposing all cryptographic identities scoped to that node. Unit 42 released an open-source tool, Spooffe, for defenders to test identity exposure, and notes the technique has not been observed exploited in the wild.

Palo Alto Unit 42·Sep 10, 10:00 UTC · 6d agoResearch1

The Pelican comparison grid for Astra is pretty interesting

Simon Willison's pelican SVG comparison shows GPT-6 Astra producing markedly better images than GPT-5.6 Sol, Terra, and Luna across reasoning levels.

Willison generated pelicans-riding-bicycles SVGs with newly accessed GPT-6 Astra at low through max reasoning levels and rendered them in a comparison grid against GPT-5.6 Sol, Terra, and Luna. Astra's outputs were markedly more coherent, while even the best GPT-5.6-Sol images remained largely abstract shapes. Astra does not support a reasoning=none setting, so all comparisons involved reasoning-enabled runs.

Simon Willison·Sep 4, 23:59 UTC · 11d agoAI research

WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data

WearableQA benchmark introduces 4,084 questions over real longitudinal wearable data, showing 14 LLMs score 19.6-72.9% on health reasoning, far from solved.

WearableQA is a benchmark of 4,084 10-option multiple-choice questions built from wearable time series, blood biomarkers, and demographics of 200 real users with up to 500 days of daily measurements. It defines 16 question types along two axes: data versus health reasoning, and single- versus cross-signal reasoning, using a dual-grounding framework combining literature and population-validated patterns. Evaluations of 14 proprietary and open-source LLMs show accuracy ranging from 19.6% to 72.9% against a 10% chance baseline, with most models below 60%.

Hugging Face daily papers·Sep 3, 20:00 UTC · 13d agoAI research

Mi-Ripple: Restoring Images Degraded by Iterative AI Editing

Mi-Ripple is a diagnosis-guided restoration workflow that removes digital ripple artifacts introduced by iterative AI image editing while preserving structure.

Iterative reference-conditioned image editing can introduce grid-like and granular textures known as digital ripple. Mi-Ripple separates periodic lattice artifacts from content-entangled granular texture, then applies selective spectral notching, structure-aware smoothing, and cleaned-reference regeneration. In fourteen notch-only executions, whole-image residual standard deviation was 0.08-0.44 in CIELAB lightness units, and reference cleaning reduced output debris density by 45% in a paired example.

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research

Bad Likert Judge: A Novel Multi-Turn Technique to Jailbreak LLMs by Misusing Their Evaluation Capability

Unit 42 details the Bad Likert Judge multi-turn jailbreak that abuses LLMs' evaluation capability, raising attack success rates over 60% across six frontier models.

Palo Alto Networks Unit 42 describes the Bad Likert Judge technique, a multi-turn jailbreak that asks a target LLM to act as a Likert-scale judge scoring the harmfulness of example responses. The highest-rated example in each scale can carry harmful content, bypassing the model's internal guardrails. Testing across six state-of-the-art text-generation LLMs showed an average attack success rate increase of more than 60% versus plain attack prompts, with tested models anonymized. The technique targets edge cases rather than typical use, and the article positions the work as guidance for defenders on potential jailbreak risks.

Palo Alto Unit 42·Aug 17, 12:34 UTC · Aug 17, 2026AI safety & security

X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation

XPeng AI's X-AuT prunes speech LLM audio encoders, cutting Qwen3-ASR-0.6B error from 5.61% to 5.27% with fewer parameters.

X-AuT is a progressive compression framework for speech LLM audio encoders that selects layer combinations via short behavioral probes and restores pruned models using cross-scale distillation and LoRA finetuning while keeping the language-model backbone frozen. Compressing Qwen3-ASR-0.6B from 18 to 16 audio-encoder layers lowered macro-average error from 5.61% to 5.27% on ten Chinese-English benchmarks. A 14-layer model reached 5.75% error with 20.7% fewer audio-tower parameters, and progressive pruning outperformed direct pruning (5.75% vs 6.73%).

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research

From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge

Interpretability study traces how Qwen, Llama, and Gemma route query information and internal knowledge across layers when answering questions.

Researchers used layerwise interventions on hidden states to separate query-routing signals from target knowledge in language models, testing Qwen, Llama, and Gemma on country-continent questions with varied answer types. A pair-conditioned request direction strengthens before interventions alter downstream knowledge, opening a causal window while answer-supporting content is still forming. Trajectories differ by model: Gemma shows a partially overlapping mid-layer routing profile, while Llama has no sustained routing-effect window under the same gates.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:39 UTC · 6d agoAI research1

"Shut Up and Let Me Enjoy My Otome": Understanding and Measuring the Toxicity in Otome Game Communities

First large-scale study finds 22.20% of Weibo otome game posts toxic versus 3.71% on Reddit, with LLM detectors reaching 0.82 F1.

Researchers present the first large-scale measurement of toxicity in otome game communities, introducing OtomeSCAN, which collected and analyzed 620,045 posts from Weibo and Reddit over 18 months. They manually annotated 4,308 posts, identified eight target groups, and evaluated seven toxicity detectors, with their best LLM-based model reaching F1-scores of 0.82 on Weibo and 0.78 on Reddit. The study found 22.20% of Weibo posts were toxic versus 3.71% on Reddit, and toxicity rose to 37.09% within 72 hours during an external attack on Weibo. The authors also flagged 191 potential-coordination clusters, 64.40% of which targeted game developers.

arXiv cs.CR·Sep 7, 21:44 UTC · 9d agoResearch

Mercator ↔ Equal Earth

Simon Willison used GPT-6 Astra (medium) in ChatGPT Work to build an animated D3 transition between Mercator and Equal Earth map projections.

Willison built an animated transition between the Mercator and Equal Earth map projections using D3. The tool was generated by GPT-6 Astra (medium) in ChatGPT Work. Equal Earth is a projection recently voted on at the UN. The post is a vibe-coding demonstration rather than a security or major model event.

Simon Willison·Sep 7, 16:24 UTC · 9d agoAI tools & infra1

Quoting Calif Research

Calif Research demos WeWorm, an AI-built zero-click worm that spreads via WeChat calls on iOS and Android without user interaction.

Calif Research released a demo of WeWorm, described as the first zero-click worm to spread through WeChat calls across iOS and Android; victims need not answer the call and hear nothing even if they do. The team used AI to find the underlying bug and write a remote code execution exploit in about two days, then spent one more week building the worm. The researchers argue AI can now do most of the exploit-development work, with humans supplying judgment on targeting and safe testing.

Simon Willison·Sep 10, 00:56 UTC · 6d agoAI safety & security

Legora reviewed 41 documents in minutes with GPT-6 Astra

Legal-tech firm Legora says GPT-6 Astra reviewed 41 financial documents in minutes, catching all four planted errors and boosting accuracy about 40%.

Legal technology company Legora reported using OpenAI's GPT-6 Astra to review 41 financial-statement documents in minutes. The workflow found all four planted errors and improved performance by nearly 40% compared to prior processes. The case study highlights AI-assisted financial review adoption in professional services.

OpenAI News·Sep 3, 12:00 UTC · 13d agoAI industry

Molecular Déjà Vu: Digit-Level Retrieval of Published Values in Frontier Language Models

Audit of 22 frontier models finds widespread verbatim retrieval of published molecular property values, with higher reasoning increasing recall of memorized numbers.

An arXiv audit tests 22 frontier LLMs across 12 molecular regression benchmarks for verbatim retrieval of published values. More than 50% of the LLMs show verbatim retrieval on five datasets, and identical experiments are flagged 89% more often at a high reasoning level than at the lowest one. Suppressing retrieval moves model prediction errors closer together in relative terms, suggesting predictive capability is not determined solely by memorized values.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:32 UTC · 12d agoAI research1

New AI Attack Hides Malicious Instructions in Normal-Looking Text to Evade Safety Filters

Check Point researchers show crafted prose hides policy-violating instructions that bypass all tested LLM gatekeepers, including GPT-4o mini and Llama Guard 3.

A new prompt-crafting technique embeds malicious payloads inside grammatical, natural-looking text without Base64, invisible Unicode, or obvious encodings, defeating lightweight pre-screening gatekeepers. In testing, all four evaluated gatekeeper models—gpt-4o-mini-2024-07-18, gpt-oss-safeguard:20b, claude-3-haiku-20240307, and llama-guard3:8b—classified the crafted wrappers as safe at a 100% bypass rate across 23 obfuscated prompts. GPT-5 Thinking in high-reasoning mode recovered and acted on the hidden instruction in 17 of 18 tests (~94.4%), often spending over a minute and multiple Python executions. Researchers recommend paraphrasing untrusted input, hardening gatekeeper policies, and applying defense-in-depth controls for agentic deployments.

GBHackers·Sep 11, 05:53 UTC · 5d agoAI safety & security 2 sources

When Should LLMs Abstain? Chain-of-Self-Questioning for Selective Risk Control

Chain-of-Self-Questioning prompting cuts LLM wrong-answer commitments 32% relative while raising answered accuracy, holding across eleven model families.

The paper introduces Chain-of-Self-Questioning (CoSQ), a prompt-only framework that makes LLM answer commitment conditional on an explicit assessment of the information required to answer. On an 817-item TruthfulQA multiple-choice set, Grounded-CoSQ at τ=0.90 reduced mean unconditional wrong-commitment rate from 13.1% under chain-of-thought to 8.9% (a 32.1% relative reduction), while raising answered accuracy from 86.9% to 89.7% at 87.6% coverage. Improvements held across eleven open-weight and hosted model families and at every evaluated threshold, with convergent evidence from a Natural Questions short-answer evaluation.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:52 UTC · 1d agoAI research

Rare Not Random Using Token Efficiency for Secrets Scanning

Researcher proposes token efficiency (string length divided by BPE token count) as a better post-regex filter than entropy for secrets scanning, validated on CredData.

The post explores whether Byte-Pair Encoding tokenization can replace Shannon entropy as the primary filter for candidate secrets captured by regex in tools like Gitleaks. It defines 'token efficiency' as string length divided by token count under the cl100k_base tokenizer; secret-like strings such as GitHub tokens tokenize into many small tokens and score low, while natural text scores high. Evaluating labeled secrets from the CredData dataset shows a usable separation, with roughly 2.5 suggested as a minimum cutoff versus Gitleaks' 3.5 entropy threshold. The technique is positioned as a post-regex filtering step rather than a standalone detector.

Lobsters · security·Sep 12, 06:42 UTC · 4d agoResearch

LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics

LexFlip releases 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving tokens, exposing weaknesses in embedding-based meaning preservation metrics.

LexFlip provides 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving 0.93 of tokens, creating dissociation items that break monotone token-overlap metric validation. The seven embedding and BERTScore metrics tested register only 0.022-0.039 of their identical-to-unrelated range on these edits, versus 0.670 for bidirectional NLI. Against FrJudge, with a measured human ceiling of r=0.597, a bare length feature outscores every semantic metric tested.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 15:49 UTC · 12d agoAI research

Retrofitting Code Using LLMs to Support Exceptional Behavior

EXCODER combines static/dynamic analysis with LLMs to retrofit exception-handling code, achieving 85.92% pass@1 with Qwen 2.5 Coder 32B on Java benchmarks.

The paper introduces the task of retrofitting existing code with Exception Related Code (throw statements, guarding conditions, try/catch blocks) so that given Exceptional Behavior Tests pass. EXCODER performs context engineering by integrating static and dynamic program analysis output with LLMs; it was evaluated on a benchmark built from 304 methods across 75 GitHub Java projects. Combined with Qwen 2.5 Coder 32B, EXCODER achieves pass@1, 5, and 10 rates of 85.92%, 86.18%, and 86.51%, roughly 13 percentage points over baseline, and manual inspection reveals remaining limitations.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:18 UTC · 7d agoAI research1

WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data

WearableQA benchmark tests LLM health reasoning over longitudinal wearable data; the best of 14 evaluated LLMs reaches 72.9% accuracy.

WearableQA comprises 4,084 ten-option multiple-choice questions built from wearable time series, blood biomarkers, and demographics of 200 real users with up to 500 days of daily measurements. It defines 16 question types along two axes: data versus health reasoning, and single- versus cross-signal reasoning. Evaluation of 14 proprietary and open-source LLMs shows performance from 19.6% to 72.9% against a 10% chance baseline, with most models below 60%.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:52 UTC · 12d agoAI research