Search: “quic”

38 stories

DomainAll Cyber AIWindow24h 3d 7d 30d AllIn the wild Important only Show duplicates

CategoryAll Malware 15 Vulnerability 13 Threat actor 8 Ransomware 7 Exploit / PoC 6 Industry 3 Policy & legal 2 Tools 2 AI tools & infra 2 Research 2 Data breach 2 AI safety & security 1 Advisory 1

SeverityAll critical 24 high 20 medium 12 low 6 info 2

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

PANORAMA: Panoptic Grounded Captioning via Mask Proposal Selection

Researchers introduce PANORAMA, a vision-language model grounding caption phrases in pixel masks, alongside the PanoCaps benchmark and gPQ metric.

The paper studies panoptic grounded captioning, requiring VLMs to describe foreground and background regions while grounding each phrase with pixel-level masks. The authors release PanoCaps, a human-annotated benchmark built from panoptic segmentation datasets, plus a phrase-mask matching protocol and a generalized Panoptic Quality (gPQ) metric. PANORAMA formulates grounding as selection from phrase-conditioned mask proposals generated by a pretrained segmenter, achieving the best overall grounding on PanoCaps and matching or exceeding specialized models. Code, data, and models are publicly available.

Hugging Face daily papersupdated Sep 16, 17:59 UTC · 20h agofirst Sep 15, 20:00 UTC · 1d agoAI research 2 sources

LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence

LimiX-2 scales Contextual Mechanism Networks pretrained via context-conditional masked modeling, beating tabular foundation models on TabArena, TALENT, and BCCO benchmarks.

LimiX-2 is a new tabular model in the LimiX family, developed through model and data scaling guided by previously established scaling laws under the Contextual Mechanism Networks (CMNs) paradigm. It is pretrained with Context-Conditional Masked Modeling (CCMM) on synthetic datasets generated by structural causal models spanning diverse graph structures, functional mechanisms, and observation processes. It outperforms dataset-specific models and tabular foundation models on TabArena, TALENT, and BCCO, and its feature attention encodes direct causal relationships, enabling accurate causal skeleton recovery.

arXiv cs.AI / cs.LG / cs.CLSep 15, 17:30 UTC · 1d agoAI research 2 sources2

DragonForce Hackers Abuse Microsoft Teams Relays to Hide Backdoor.Turn C2 Traffic

Symantec and Carbon Black found DragonForce-linked hackers using Backdoor.Turn, a Go RAT tunneling C2 through Microsoft Teams TURN relays via QUIC.

Broadcom's Symantec and Carbon Black documented Backdoor.Turn, a Go-based RAT associated with the DragonForce ransomware group, deployed at a major U.S. services firm where attackers remained on the network for one to two months. The backdoor obtains an anonymous Teams visitor token from Microsoft's Skype-backed identity services, uses a legitimate Microsoft TURN relay for connection setup, and runs a QUIC session to the attacker's C2 server, so defenders see only outbound traffic to legitimate Teams servers. Initial access is suspected to involve exploiting a SQL/MS-SQL vulnerability or an access broker, followed by a PowerShell-dropped ZIP, DLL side-loading, and BYOVD evasion using a Huawei driver; the RAT is injected into DbgView64.exe after DragonForce ransomware deployment for continued access. First publicly documented TURN abuse by the group, it leverages Praetorian's Ghost Calls technique and supports command execution, network scanning, LDAP/AD search, lateral movement, and browser credential theft.

The Hacker NewsAug 18, 12:13 UTC · Aug 18, 2026Malware CVE-2023-52271 CVE-2025-61155 CVE-2025-10551

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.CLSep 4, 17:32 UTC · 12d agoAI research1

Microsoft Patches 398 Flaws Including a Windows Driver Zero

Microsoft's August Patch Tuesday fixes 398 CVEs, including actively exploited Windows zero-day CVE-2026-68820 used by Lazarus for SYSTEM privilege escalation.

Microsoft's August 2026 Patch Tuesday release addresses 398 new vulnerabilities, 62 rated Critical by the Zero Day Initiative, with CVE-2026-68820 (CVSS 7.0), a use-after-free in the afd.sys Windows kernel driver, the only flaw flagged as actively exploited. Check Point Research attributed exploitation of the privilege escalation bug to Lazarus' Operation Dream Job campaign. Four unauthenticated remote code execution flaws rated 9.8 affect Windows DNS Server (CVE-2026-62878, called wormable by ZDI), Windows Deployment Services (CVE-2026-62893), Microsoft QUIC (CVE-2026-62815) and HPC Pack (CVE-2026-59124). The release also completes a two-part SharePoint fix with RCE component CVE-2026-63520, following July's authentication bypass fix CVE-2026-55040.

The Hacker NewsAug 11, 20:10 UTC · Aug 11, 2026Exploit / PoC in the wildCVE-2026-68820 CVE-2026-62878 CVE-2026-62893+4 CVEs

Claude's new system prompt really doesn't want to reproduce song lyrics

Anthropic published updated Claude consumer system prompts, including changes steering the model away from reproducing song lyrics, likely over copyright concerns.

Anthropic publishes system prompts for Claude.ai and Claude mobile apps, including historic revisions, and has reorganized them into an index with per-model pages such as the Haiku 4.5 page showing the original October 15, 2025 prompt and an updated January 18, 2026 version. The latest consumer prompt strongly discourages reproducing song lyrics, a behavioral constraint likely tied to copyright considerations. Prompts for Claude Cowork and Claude Code are not included in the published set.

Simon WillisonSep 2, 14:16 UTC · 14d agoAI safety & security1

CDN Tsunami Attack Abuses HTTP/3 Translation for Up to 350x DoS Amplification

Researchers disclosed two DoS attacks exploiting CDNs' HTTP/3-to-HTTP/1.1 translation to amplify low-bandwidth requests up to 350x against origin servers.

Researchers named the two techniques HTTP/3 Bandwidth Amplification (HBA) and HTTP/3 Connection Amplification (HCA), exploiting the gap where CDNs speak HTTP/3 to browsers but only HTTP/1.1 to origins. All six tested CDNs (Alibaba, Baidu, Cloudflare, CloudFront, Fastly, Tencent) were susceptible to HBA, and five to HCA, with Cloudflare immune to HCA because it buffers full requests before opening backend connections. Attackers using under 5 Mbps forced over 100 Mbps at origins, and the 350x factor applies only to Alibaba, Baidu, and Tencent, which support the QPACK dynamic table. Testing identified 42,330 potentially vulnerable subdomains out of 151,685 hosted on the six providers across the Tranco Top 1M.

The Hacker NewsAug 20, 11:39 UTC · 28d agoResearch

Structural Jailbreaks Generalize but Do Not Compound: A cross-provider and multilingual study of Involuntary In-Context Learning

Researchers show IICL structural jailbreaks generalize to Google Gemini, lifting attack success to 80-100% on harm and financial benchmarks; non-English prompts attenuate it.

The study red-teams two Google Gemini models with Involuntary In-Context Learning (IICL), a structural jailbreak reframing harmful requests as the final cell of a data-labeling task. IICL lifts attack success from at most 6.7% to 80-90% on HarmBench and 97-100% on financial abuse (FinProof), an order of magnitude above prior results on OpenAI's GPT-5.4. Against a compounding hypothesis, forcing IICL output into Spanish, Hindi, or Arabic attenuates the attack in 11 of 12 conditions, attributed to a 'relevance curse' producing lower-quality harmful content in lower-resource languages. Findings replicate under an independent non-Google judge (Cohen's kappa 0.86 over 377 paired verdicts).

arXiv cs.CRSep 8, 07:38 UTC · 9d agoAI safety & security1

New KATARU IoT Malware Packs Linux Privilege Escalation Exploits and Mirai-Style DDoS Attacks

Nozomi Networks identified KATARU, a new Mirai-style IoT botnet delivered via Telnet brute force that uses Linux privilege-escalation exploits and encrypted C2 for DDoS floods.

Nozomi Networks identified KATARU in August after a Telnet password-guessing attack against a honeypot retrieved an ARM payload. The malware attempts exploits for CVE-2026-46300 (Fragnesia), CVE-2026-43284 (DirtyFrag), and CVE-2026-31431 (Copy Fail), plus a cgroup v1 release_agent escape, and persists via systemd services, cron tasks, rc scripts, OpenWrt hooks, and Android boot locations. Its C2 uses X25519 key exchange with ChaCha20-Poly1305 encryption and supports TCP, UDP, ICMP, HTTP, QUIC, and DNS floods, plus SSH brute forcing and command execution; embedded exploit shellcode in the ARM build targeted x86, suggesting untested copied code.

Cyber Security NewsSep 11, 14:17 UTC · 5d agoMalware in the wild 2 sources CVE-2026-46300 CVE-2026-43284 CVE-2026-31431

Right Tool, Right Job: Native-Language Evaluation, Tokenizer Sensitivity, and Methodological Findings from a French-Only BabyLM

French BabyLM entry METRON-FR (125M GPT-2, 92.47M words) shows tokenizer artifacts dominate child-scale zero-shot evaluation; proposes standard diagnostics.

METRON-FR is a 125M-parameter GPT-2 pretrained on 92.47M French words, submitted to the BabyLM 2026 Strict track, scoring 85.97% on the native Quebec-French QFrBLiMP benchmark and 62.80% on the BabyLM-weighted leaderboard. A cross-lingual GLUE protocol combining French task-data translation with rank-16 LoRA shows relational tasks gain while world-knowledge tasks regress. Bilingual Lexicon Induction reaches p@1 of 68.84%, 18x above chance, and ablations show single-token zero-shot scoring is dominated by tokenizer and template artifacts at child scale.

arXiv cs.AI / cs.LG / cs.CLSep 15, 16:43 UTC · 1d agoAI research

BlackHatSect0r Hackers Disable AI Safety Controls to Automate Credential Theft and Cyberattacks

French-speaking crew BlackHatSect0r disabled AI agent safety controls to automate scanning, credential harvesting, and vishing, exposing 16,834 stolen credentials.

Socradar researchers analyzed the exposed infrastructure of a French-speaking crew called BlackHatSect0r && DXQRTXX, which ran a Nous Research Hermes agent on a DeepSeek model with safety controls removed via HERMES_DISABLE_SAFETY=1. A custom Go-based C2 platform, DXSCAN, was exposed on port 8080 with over 200 secret-detection patterns, a vault of 16,834 harvested credentials, and scanning activity queuing 2.75 million domains and reaching more than 726,000 hosts. The kit also held a database of roughly 450,000 French telecom subscriber records used to prepare vishing lures impersonating Société Générale, plus JWT-forging tooling for a cryptocurrency exchange. Most confirmed compromises relied on exposed secrets and cloud misconfiguration rather than novel exploits; the one cited vulnerability, CVE-2026-42530, is an NGINX HTTP/3 QPACK use-after-free fixed in version 1.31.2.

GBHackersupdated 09:59 UTC · 4h agofirst 07:00 UTC · 7h agoThreat actor in the wild 4 sources CVE-2026-42530

Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering

Evaluation of twelve LLMs on 222 clinical questions shows verbatim quotes rarely substantiate claims; claude-opus-5 fully substantiates only 37.1%.

The authors build a standardized harness over four clinical practice guidelines and evaluate twelve LLMs on 222 synthetic clinical questions, measuring citation attachment, verbatim quote production, and claim substantiation. Most models attach verbatim quotes to over 90% of claims from prompting alone, though lightweight models like claude-haiku-4.5 struggle. Quotes frequently fail to substantiate claims: claude-opus-5 quotes 98.0% of claims but fully substantiates only 37.1%, exposing a capability gap for verifiable clinical QA.

arXiv cs.AI / cs.LG / cs.CLSep 14, 17:53 UTC · 2d agoAI research

Operation QUICSILVER Targets Myanmar Government and IT with QUICAgent Backdoor

Seqrite Labs details Operation QUICSILVER, a China-nexus espionage campaign targeting Myanmar government and IT with graduation-invite lures deploying QUICAgent backdoor.

Seqrite Labs reported Operation QUICSILVER, a cyber espionage campaign against Myanmar's government and IT sectors attributed with moderate confidence to a China-nexus actor. Since April 2026, attacks used fake Belgian-Myanmar holiday and Burmese graduation ceremony invitation lures delivered via VHD/LNK files that abuse ftp.exe (LOLBAS) to assemble and launch QUICAgent, a Go backdoor communicating over QUIC on UDP 443 with five commands and Startup-folder persistence. Separately, China-linked Mustang Panda was observed using an updated COOLCLIENT backdoor with a signed kernel-mode driver across Myanmar, Mongolia, Pakistan, and Russia.

The Hacker NewsAug 24, 11:51 UTC · 24d agoThreat actor in the wild1

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.CLSep 4, 15:49 UTC · 12d agoAI research

Critical Unbound DNSSEC Validator Flaw Could Allow RCE via a Malicious DNS Zonenew

NLnet Labs patched a critical heap overflow (CVE-2026-81642) in Unbound's DNSSEC validator that a malicious DNS zone can trigger for possible RCE.

Unbound 1.26.1, released September 17, fixes nine CVEs, including critical heap overflow CVE-2026-81642 (CVSS 9.1) in the DNSSEC validator, which affects every version through 1.26.0 and can be triggered by an attacker controlling a malicious zone that serves a DNSKEY record with a compression pointer into its own data, causing denial of service and possibly remote code execution. CVE-2026-82717, a heap corruption in CNAME synthesis reported by Ben Morris of Anthropic, could also lead to RCE under certain systems and compilation options. NLnet Labs has not reported exploitation and CISA marks CVE-2026-81642 exploitation as none; standalone and combined patches are available alongside the 1.26.1 release.

The Hacker News12:30 UTC · 1h agoVulnerability CVE-2026-81642 CVE-2026-82717 CVE-2026-81634+7 CVEs

Testing race conditions with memory access tracing and stack-based delay injection

Google Project Zero released MAccConc, Linux kernel tooling that traces memory accesses to explore and test race condition interleavings.

A Google Project Zero researcher published MAccConc (Memory Access Concurrency), tooling for exploring possible interleavings of multithreaded test cases in the Linux kernel, available on GitHub. The tools use KCOV with ASAN outline-mode instrumentation to record per-access memory traces, enabling automatic testing of all A-B-A interleavings plus terminal and GUI explorers for manual analysis. The work targets confirming race condition candidates, building reliable regression tests, and enabling concurrency fuzzing, drawing on ideas from SKI and Ned Williamson's sockfuzzer.

Google Project ZeroSep 8, 07:00 UTC · 9d agoResearch2

TWINLOOT Abuses SharePoint and Teams to Steal Credentials and Move Across Networks

Ontinue disclosed TWINLOOT, a Python implant hiding C2 in SharePoint dead drops and Teams TURN relays, harvesting credentials and pivoting via reverse SOCKS5.

Ontinue's Cyber Defense Center identified TWINLOOT during a July 2026 campaign investigation: a modular, PyArmor-hardened Python implant (a 39 MB bootstrap-fat.pyc loader) whose entire C2 infrastructure lives inside trusted Microsoft services. Tasking flows through SharePoint Online file dead drops polled every 15 seconds via the Microsoft Graph API, while interactive operator access uses WebRTC DataChannels relayed by Microsoft Teams TURN servers; Graph traffic is driven by the victim's own headless Edge browser. The implant harvests Windows credentials with fake lock screens, offers a reverse SOCKS5 pivot for lateral movement to SMB, RDP, WinRM, and MSSQL, executes commands, and persists on hosts. Initial access is assessed to be Teams social engineering masquerading as IT support, prompting a PowerShell command to download the payload.

The Hacker NewsAug 18, 15:17 UTC · 29d agoMalware1

Rosetta at AlexandriaX-2026: LoRA-Adapted NileChat for Context-Aware Dialectal Arabic Dialogue Translation

Rosetta ranks 4th and 5th in AlexandriaX-2026 dialectal Arabic dialogue translation using a LoRA adapter on NileChat-3B, finding limited pretraining benefit.

The Rosetta system for the AlexandriaX-2026 shared task fine-tunes a LoRA adapter on NileChat-3B for context-aware English-to-dialectal Arabic dialogue translation. The adapter was additionally pretrained on MADAR and PADIC dialect corpora for the unconstrained track. It achieved spBLEU 26.10 (4th, constrained) and 25.09 (5th, unconstrained). External dialect pretraining improved only two of thirteen dialects while slightly degrading overall performance, indicating negative transfer.

arXiv cs.AI / cs.LG / cs.CLSep 9, 16:15 UTC · 7d agoAI research1

Microsoft Patch Tuesday for August 2026 Fixed a Zero

Microsoft's August 2026 Patch Tuesday fixes 398 CVEs, including an actively exploited WinSock zero-day CVE-2026-68820 and wormable DNS RCE CVE-2026-62878.

Microsoft's August 2026 Patch Tuesday addresses 398 CVEs across Windows, Office, Azure, Exchange Server, SharePoint, Teams, .NET and other products, with 62 rated Critical. The actively exploited CVE-2026-68820 is a use-after-free in the Windows WinSock afd.sys driver that can yield SYSTEM privileges via a race condition. CVE-2026-62878 is a critical, wormable stack buffer overflow in Windows DNS Server allowing unauthenticated, remote code execution without user interaction. Other notable fixes include WDS TFTP RCE CVE-2026-62893, a QUIC flaw affecting roughly 13.5 million sites (CVE-2026-62815), an Exchange authentication bypass EoP demonstrated at Pwn2Own Berlin (CVE-2026-62911), and two TPM 2.0 reference implementation fixes.

Security AffairsAug 12, 06:49 UTC · Aug 12, 2026Advisory in the wildCVE-2026-68820 CVE-2026-62878 CVE-2026-62893+5 CVEs

Towards a Deterministic Math Solver for Clinical Language Models

Paper shows handing arithmetic to a deterministic Python solver beats direct model calculation at 32B but not reliably at 7B on MedCalc-Bench.

Researchers test a Program-Solve interface where clinical LLMs write case-specific Python executed by a restricted local solver instead of doing arithmetic directly. On MedCalc-Bench Verified (1,100 cases, 55 calculators), Qwen2.5-32B-AWQ scored 90.53% with solver handoff versus 83.47% with direct arithmetic (+7.05 points), while Qwen2.5-7B gained an unreliable +3.29 points with a confidence interval spanning zero. The authors audited the benchmark against clinical guidelines and flagged 16 of 55 calculators for version, use, or coefficient concerns.

Hugging Face daily papersSep 8, 20:00 UTC · 8d agoAI research1

Microsoft patches 400+ vulnerabilities, one zero-day under attack (CVE-2026-68820)

Microsoft's August 2026 Patch Tuesday fixes 400+ vulnerabilities, including an actively exploited Windows zero-day (CVE-2026-68820) used by North Korean attackers.

Microsoft's August 2026 Patch Tuesday fixes over 400 vulnerabilities, including CVE-2026-68820, a use-after-free in the Windows Ancillary Function Driver exploited in the wild by North Korean actors deploying a kernel-mode rootkit in Operation Dream Job. Critical unauthenticated remote code execution flaws in Microsoft QUIC (CVE-2026-62815) and Windows DNS (CVE-2026-62878) were also patched, alongside a SharePoint RCE chain combining CVE-2026-63520 with CVE-2026-55040. Researcher Nightmare-Eclipse released ShieldBreak, a PoC bypassing the July RoguePlanet Microsoft Defender patch (CVE-2026-50656), confirmed working by Will Dormann on Windows 11.

Help Net SecurityAug 12, 00:00 UTC · Aug 12, 2026Exploit / PoC in the wildCVE-2026-68820 CVE-2026-62832 CVE-2026-72971+6 CVEs1

27.5KB language-agnostic WebGPU syntax highlighter

A developer released gpu-lexer, a 27.5KB language-agnostic syntax highlighter that uses a tiny WebGPU model to label code tokens in the browser.

gpu-lexer splits source into words, whitespace, and symbols, then a small WebGPU model uses local and whole-file context to assign nine token classes, working on languages never seen in training. On held-out files, 12.57% of token labels differ from Shiki, though this measures agreement with Shiki rather than objective correctness. In benchmarks against Shiki 4.4.3, Prism.js, Highlight.js, Sugar High, and Starry Night, it highlighted 10 concatenated copies of three.min.js (5.56M characters) about 10x faster on an Apple M4 Pro in Chrome 152. The author frames it as an experiment, not a grammar-equivalent highlighter.

Hacker News · securitySep 9, 01:14 UTC · 8d agoAI tools & infra HN 21↑ · 5 comments1

IndicTriMix: Developing Language Identification Datasets and Models for Tri-Language Code-Mixing

Researchers release IndicTriMix benchmarks and fine-tuned MuRIL and XLM-RoBERTa models for token-level language identification in tri-language code-mixed text.

The paper formulates token-level language identification in code-mixed text as a sequence labeling task and fine-tunes MuRIL and XLM-RoBERTa transformer models for Indian languages. It evaluates on Hindi, Gujarati, and Bengali configurations with manually annotated test sets and proposes two code-mixed generation approaches using parallel trilingual sentences. A public benchmark, annotated test sets, and fine-tuned models are released for reproducibility.

arXiv cs.AI / cs.LG / cs.CLSep 10, 17:36 UTC · 6d agoAI research1

Convergent Emergence of In-Context Learning Across Modalities

Controlled experiments show few-shot in-context learning emerges across six modalities including language, genomes, images, and proteins, partially supporting a convergence hypothesis.

The paper tests the Convergent Emergence Hypothesis: that few-shot in-context learning, when it emerges, shares a common cross-modality difficulty profile. A controlled framework instantiated the same task suite across six modalities: language, genome, integer sequences, time series, images, and proteins. Paired-mapping ICL emerged in all six modalities, surpassed controlled baselines, and showed correlated per-task effects in five of them, providing partial support for the hypothesis.

Hugging Face daily papersSep 11, 20:00 UTC · 5d 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.CLSep 10, 17:39 UTC · 6d agoAI research1

ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs

ModaLens image-swap audit shows report availability cuts MedGemma-27B image sensitivity on MIMIC-CXR from 20.94% to 4.26% answer changes.

ModaLens is a paired image-swap audit measuring how report availability affects image sensitivity in report-conditioned medical VLMs. On MedGemma-27B across 3,199 paired MIMIC-CXR cases from 293 patients (14 questions per case), generated answers changed on 4.26% of image-swap trials with the report versus 20.94% without it, a 16.7-point paired difference (95% CI 15.6-17.7). The original prompt with a lowercase first-token readout gave 4.70% versus 17.07%, and the direction replicated in two further model lineages. Labels derived from reports limit conclusions about visual correctness; code, prompts, and run records are publicly released.

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

Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models

New ROBORMBENCH benchmark shows vision-language reward models can flip robot success/failure judgments when goal instructions are paraphrased.

The authors show that paraphrasing the instruction alone can substantially change progress scores from VLM reward models, even flipping identical robot trajectories between failure and success. ROBORMBENCH comprises 2,390 real-robot trajectories with ground-truth progress labels and 21,673 verified paraphrases covering lexical, syntactic, and action-goal rewrites. Instability is widespread across proprietary and open-source VLMs, grows with more divergent rewrites, and is not reliably reduced by scale or explicit reasoning, while trajectory-grounded dedicated reward models are markedly more stable.

arXiv cs.AI / cs.LG / cs.CLSep 4, 17:47 UTC · 12d agoAI research1

Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models

Researchers introduce KoNA, a benchmark exposing vision-language models' failures at selective non-compliance, plus fine-tuning that improves refusal and abstention accuracy.

KoNA is a benchmark for evaluating selective non-compliance in vision-language models across five categories: False Premise, Visual Inaccessibility, Universal Unknown, Task Feasibility and Safety. It tests both query-level and component-level non-compliance using paired single and compound queries, and evaluations across diverse VLMs show models often fail to refuse, correct or abstain appropriately, with failures worsening on compound queries. Fine-tuning VLMs on KoNA examples substantially improves non-compliance accuracy while largely maintaining performance on fully answerable tasks.

Hugging Face daily papersSep 3, 20:00 UTC · 13d agoAI research1

LLM Agents as Computational Typologists

AUTOTYPOLOGIST is an LLM agent that performs evidence-grounded linguistic typology analysis over 25 open-source reference grammars.

The agent retrieves relevant grammar sections, analyzes interlinear glossed text (IGT), and iteratively reasons over typological hypotheses in a ReAct-style workflow. It was evaluated on typological feature coding against expert annotations and hypothesis testing against universals using 25 open-source reference grammars. Results suggest LLM agents can support scalable, inspectable crosslinguistic analysis but still require expert validation.

arXiv cs.AI / cs.LG / cs.CLSep 7, 17:35 UTC · 9d agoAI research1

Claude is a Contrarian

Opinion piece argues Claude habitually contradicts explicit user instructions, injecting contrarian content despite CLAUDE.md rules and user objections.

A developer recounts repeated instruction-following failures with Claude, claiming it contradicts explicit requests, adds unnecessary work, and ignores AGENTS.md and CLAUDE.md directives. The author contrasts this with OpenAI, DeepSeek, and Qwen models, which he says more readily apologize and undo mistakes. He theorizes Claude's training makes it assume the human is wrong and needs correcting. The post is personal commentary with no benchmarks or systematic evaluation.

Hacker News · AISep 14, 16:14 UTC · 2d agoAI industry HN 54↑ · 48 comments

Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning

Tiny Aya L2-Thinker, a 3.35B model, achieves over 93 percent in-language reasoning across 60 languages via optimized multilingual data mixing; weights released.

The paper studies L2 reasoning, the ability to reason consistently in the language of the user's prompt, approached through SFT data composition and scheduling. Tiny Aya L2-Thinker (3.35B) achieves an in-language reasoning rate above 93 percent across 60 languages on six benchmarks spanning math, commonsense, instruction following, open-ended generation, and cultural reasoning. Findings show generalization to held-out languages comes from broader language coverage, multilingual non-reasoning data, and a strong English reasoning backbone, suggesting reasoning is language-agnostic and transferable without per-language supervision. Model weights and multilingual reasoning data are publicly released.

arXiv cs.AI / cs.LG / cs.CLSep 9, 16:56 UTC · 7d agoAI research1

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.CLSep 4, 16:38 UTC · 12d agoAI research

I wrote an AI textbook — how long until AI can do it better?

AI researcher Nathan Lambert argues LLMs remain weak at long-form technical writing, questioning whether models can autonomously organize scientific knowledge for breakthroughs.

Nathan Lambert describes writing a post-training textbook, Reinforcement Learning from Human Feedback, and finds today's LLMs weak at organizing long-form technical content despite becoming superhuman at coding and math. He notes GPT 5.5 Pro found deep typos across a 200-300 page manuscript while Claude models proved more useful as editors. He argues that compressing knowledge through writing is a prerequisite for autonomous scientific insight and tempers expectations for near-term AI-driven open science.

InterconnectsAug 12, 13:01 UTC · Aug 12, 2026AI research

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.CLSep 4, 17:08 UTC · 12d agoAI research1

Edge0/Edge0-35B-A3B-preview — new model trending #30 on Hugging Face

Edge0 released a 35B sparse MoE model running in under 3 GiB of memory at 15 tok/s via SSD expert offload and int4 quantization.

Edge0-35b-a3b-preview is a 35B-parameter MoE (256 experts, 4 active per token) built on Qwen3.5-MoE 35B-A3B, shipped as a 4-bit checkpoint with LoRA and prerouter adapters under Apache 2.0. The edge0 framework streams expert weights from SSD on demand, bounding peak active memory at 2.9 GiB and achieving 14.9-17.7 tok/s decode on a Mac mini M4 Pro (MLX backend). Recover-LoRA distillation keeps the int4 model within 3.9 points of its fp16 base (79.2 vs 83.2 average on OpenCompass benchmarks including AIME 2026, HumanEval, GPQA-Diamond, MMLU-Pro, and IFBench).

Hugging Face trending modelsSep 8, 13:56 UTC · 9d agoModel release

Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic

Multiverse Computing's Hugging Face post argues language models should refuse only the relevant subset of a topic instead of over-refusing whole subjects.

A Hugging Face blog post by Multiverse Computing examines refusal granularity in language models, arguing models should refuse the relevant subset of a topic rather than the entire topic. No full article text was available for additional technical detail.

Hugging Face BlogSep 8, 14:23 UTC · 8d agoAI safety & security

New infosec products of the week: September 4, 2026

Weekly roundup covers F5's AI-powered WAF enhancements, Ping Identity's personal AI agent access, Superna 2.15 cyberstorage, and BugBase Pentest Copilot Enterprise.

This week's product roundup features releases from BugBase, F5 Networks, Ping Identity, and Superna. F5's WAF for Distributed Cloud adds anomaly detection and agentic threat intelligence for real-time virtual patching, Ping Identity launched Enterprise Personal Agent Access for securing personal AI agents, Superna 2.15 adds guided event-closing workflows for cyberstorage operations, and BugBase's Pentest Copilot Enterprise automates black-box pentesting across 100 vulnerability types using real Chromium browsers.

Help Net SecuritySep 4, 00:00 UTC · 13d agoTools

Enhancing Accessibility of Medical Texts through Large Language Model-Driven Plain Language Adaptation

Study shows LLMs with Mixture-of-Agents and QLoRA finetuning effectively simplify medical texts into plain language while preserving content.

The paper evaluates Plain Language Adaptation (PLA) using GPT-4o-mini, Gemini-1.5-pro, and LLaMA in zero-shot and few-shot settings. It compares prompting strategies, QLoRA finetuning across models, and integrates Mixture-of-Agents (MoA) techniques for robustness. Results demonstrate LLM-driven PLA makes healthcare texts more comprehensible while retaining essential content.

arXiv cs.AI / cs.LG / cs.CLSep 15, 16:33 UTC · 1d agoAI research