Search: “Defused”

CVE-2026-9586 | Sangoma Switchvox Unauthenticated SQL Injection Remote Code Execution Vulnerability

CVE-2026-9586, a critical unauthenticated SQL injection to RCE in Sangoma Switchvox, is being actively exploited against internet-facing systems per honeypot observations.

CVE-2026-9586 (CVSS 4.0: 9.3) is an unauthenticated SQL injection in the /pa endpoint of Sangoma Switchvox SMB Edition, where the unsanitized PhoneIP value in XML phone notification requests reaches the PostgreSQL database. Horizon3 demonstrated escalation to OS command execution, and honeypots run with Defused Cyber observed exploitation attempts on August 30, 2026, using PostgreSQL commands with nc and curl. Fixed in Switchvox 8.4.0.2 released July 14, 2026; one attacker IP (176.65.148.184) was documented.

Horizon3.ai·Sep 1, 15:32 UTC · 15d agoExploit / PoC in the wildCVE-2026-9586

Do Input-Level Defenses Transfer to Observation-Level Attacks on VideoLLMs?

A systematic study shows input-level adversarial defenses provide inconsistent, often near-zero protection against observation-level attacks on video LLMs.

Researchers introduce DefTEval, a controlled framework testing eleven input-level defenses against five attack types across five video LLMs. Harmful-content detection rates are frequently near zero, and defenses fail even when attacks embed harmful signals in every sampled frame. Token compression discards localized safety features and modality fusion down-weights weakened visual signals, with defense outcomes dominated by model architecture rather than the defense method.

arXiv cs.CR·Sep 8, 06:59 UTC · 8d agoAI safety & security

Risky Bulletin: The EU publishes its upcoming cybersecurity standards

ETSI releases 17 draft cybersecurity standards vendors must meet when the EU Cyber Resilience Act takes effect in December 2027.

The European Telecommunications Standards Institute published 17 interim draft standards covering operating systems, routers, firewalls, VPNs, SIEMs, browsers, password managers, smart home devices, toys and wearables. They mandate basic security features such as post-sale updates, shipped SBOMs, modern cryptography and secure-by-default settings; public comments run until November, with final versions expected in December, one year before CRA compliance begins in December 2027. The newsletter also reports Irregular taking responsibility for AI test-environment escapes involving Anthropic and Meta frontier models, a breach at France's tax agency exposing 678,000+ citizens' data claimed by hacker ZeroBytes, and Kazakhstan eGov data covering 15 million citizens listed for sale on an underground forum. Additional briefs cover a $3.2 million Harmony Protocol theft crashing the ONE token 40%, Columbus Police still restoring systems two years after ransomware, DDoS attacks on Threema's provider, and Ukraine's GUR claiming a cyberattack on Wildberries.

Risky Business News·Aug 17, 05:03 UTC · Aug 17, 2026Policy & legal2

ECHO: A Matched-Contrast Benchmark for Context-Sensitive Turn-Taking in Full-Duplex Dialogue

ECHO benchmark pairs identical-overlap Chinese dialogue examples with contrasting contexts, revealing most full-duplex systems bias toward yielding the floor.

ECHO is a paired diagnostic benchmark for Chinese full-duplex turn-taking that matches examples with identical overlap transcripts but contrasting preceding multi-turn contexts, one requiring Yield and the other Keep. It also includes off-talk examples for diagnosing unnecessary yielding and introduces pair accuracy, which grants no credit to constant-action policies. Experiments across multiple full-duplex systems show most exhibit a pronounced Yield bias, performing substantially better on interruptions than backchannels. The benchmark and metadata will be publicly released.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 15:58 UTC · 1d agoAI research

SAP Commerce Cloud CVE-2026

Attackers actively exploit unauthenticated RCE CVE-2026-58231 (CVSS 10.0) in SAP Commerce Cloud days after patch release.

CVE-2026-58231, a maximum-severity flaw in SAP Commerce Cloud (CVSS 10.0), allows unauthenticated attackers to abuse a default authentication client and bypass insufficient validation, achieving arbitrary code execution and compromise of internal components. Defused Cyber researchers observed exploitation attempts against honeypots only three days after SAP released the patch. The flaw had no public PoC and was not previously known to be exploited. The attackers are unknown, though past critical SAP flaws have been exploited by China-linked groups UNC5221 and UNC5174 and ransomware gangs.

Security Affairs·Aug 15, 17:14 UTC · Aug 15, 2026Exploit / PoC in the wildCVE-2026-58231

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

SharePoint CVE-2026-55040 Comes Under Attack Following Public Exploit

Attackers are actively exploiting critical SharePoint authentication bypass CVE-2026-55040 (CVSS 9.1) days after Rapid7 released a public PoC.

CVE-2026-55040 is a critical authentication bypass (CVSS 9.1) in SharePoint Server Subscription Edition that lets unauthenticated attackers forge JWT tokens and impersonate any user, including administrators. Microsoft patched it in July 2026 Patch Tuesday, but Rapid7 published a technical write-up and Python PoC on GitHub on August 12, and Defused quickly observed attackers running the PoC against SharePoint honeypots. KEVIntel recorded 12 exploitation attempts since July 19, with eight occurring on August 12-13 from IP addresses in Hong Kong, Japan, the Netherlands, Taiwan and the U.S. Admin-level SharePoint access exposes documents and data modification and can serve as a foothold into broader Microsoft 365 infrastructure; unpatched instances should apply the July update immediately.

Security Affairs·Aug 13, 08:36 UTC · Aug 13, 2026Exploit / PoC in the wildCVE-2026-55040

What Breaks Under Pruning in Smart Homes, and When? Evaluating LLM Degradation Across Architectures and Task Complexity

Pruning study across four LLM architectures finds dense models degrade sharply on smart-home tool calling while MoE models tolerate far more.

Researchers systematically study pruning-induced degradation in smart-home tool calling across four LLMs spanning dense Transformer, dense hybrid, and mixture-of-experts architectures, combining depth, width, hybrid, and expert pruning methods, and evaluate over 19,500 instances from three datasets after post-pruning supervised fine-tuning. Dense models show narrow safe pruning regions followed by sharp degradation, while MoE models tolerate substantially more pruning. Pruning degrades grounded specificity (operation, device, argument, value) before schema-level intent, and aggressive dense pruning can induce systematic over-refusal.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:51 UTC · 1d agoAI research

Exploitation of Sangoma Switchvox flaw is underway (CVE-2026-9586)

Attackers actively exploit SQL injection flaw CVE-2026-9586 in unauthenticated Sangoma Switchvox endpoints, dropping reverse shells and second-stage cryptominer malware.

CVE-2026-9586, an unauthenticated SQL injection in Sangoma Switchvox SMB Edition 8.3, lets crafted HTTP POST requests execute arbitrary SQL against the backend PostgreSQL database. Horizon3 honeypots first saw exploitation on August 30, 2026 from IP 176.65.148.184, and dozens of additional source IPs have since joined with scanning payloads and second-stage malware that appears to be a cryptominer. The flaw was patched in Switchvox 8.4.0.2 on July 14, 2026; roughly 4,000 exposed instances exist, mostly in the United States, and exploitation is likely against most of them.

Help Net Security·Sep 2, 00:00 UTC · 14d agoExploit / PoC in the wildCVE-2026-9586

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

Three Critical VMware Flaws Allow Auth Bypass, Code Execution, and VM Escape

Broadcom patches three critical VMware flaws, including vCenter auth bypass CVE-2026-59309 and an ESX VM escape, urging immediate updates.

Broadcom issued emergency updates for VMware ESX, vCenter, Workstation and Fusion, fixing five flaws including CVE-2026-59309 (CVSS 9.8 authentication bypass in vCenter) and CVE-2026-59310 (CVSS 9.8 directory traversal allowing code execution). CVE-2026-47876 (CVSS 9.3) is an out-of-bounds write in the ESX VMXNET3 adapter that permits a virtual machine escape to the host. Broadcom found no evidence of exploitation, but Defused Cyber reported a spike in vCenter scanning on August 11, 2026 consistent with exploitation attempts against CVE-2026-59309 (VMSA-2026-0006). No workarounds are available.

The Hacker News·Aug 11, 15:42 UTC · Aug 11, 2026VulnerabilityCVE-2026-59309CVE-2026-59310CVE-2026-47876+2 CVEs1

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 Blog·Sep 8, 14:23 UTC · 8d agoAI safety & security

Three CVSS 10.0 ServiceNow Flaws Could Let Unauthenticated Attackers Execute Code and SQL

ServiceNow patched four AI Platform flaws, including three pre-authentication CVSS 10.0 issues enabling unauthenticated code execution, SQL injection, and privilege escalation.

ServiceNow released patches on August 27, 2026 for four AI Platform flaws: CVE-2026-18885 (code injection in the GraphQL Composite Data API), CVE-2026-18886 (improper access control enabling privilege escalation), and CVE-2026-74820 (SQL injection), all self-rated CVSS 10.0 and exploitable without authentication, plus CVE-2026-6876, an 8.7 sandbox escape. Updates were deployed to hosted instances, but self-hosted customers must patch affected Xanadu, Yokohama, Zurich, and Australia release lines themselves. ServiceNow says it is not aware of exploitation of the new flaws, and no public exploit code existed as of August 28, 2026; separately, Defused reported in-the-wild exploitation of the earlier CVE-2026-6875 (CVSS 9.5), later noting the captured payload matched Searchlight Cyber's PoC.

The Hacker News·Aug 28, 11:20 UTC · 19d agoVulnerability in the wildCVE-2026-18885CVE-2026-18886CVE-2026-74820+2 CVEs

One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation

A review paper frames on-policy self-distillation collapse as governed by three levers: token weighting, privileged information, and guidance decay.

The paper critically reviews On-Policy Self-Distillation (OPSD), where a language model trains on its own generations scored token-by-token by a teacher conditioned on privileged information such as reference solutions or environment feedback. It identifies collapse, the progressive narrowing of producible reasoning paths, as the dominant failure mode and analyzes it through three levers: signal weighting, the nature of privileged information, and teacher dynamics. The review is restricted to mathematical reasoning, reports no new experiments, and offers a shared vocabulary separating settled findings from disputed ones.

Hugging Face daily papers·Aug 25, 20:00 UTC · 22d agoAI research

Attackers Exploit SharePoint Authentication Bypass After Public PoC Release

Attackers exploit SharePoint authentication bypass CVE-2026-55040 (CVSS 9.1) after Rapid7's PoC release, with 12 exploitation attempts recorded since July 19, 2026.

Microsoft patched CVE-2026-55040 in its July 2026 Patch Tuesday; the 9.1-severity flaw chains four weaknesses in SharePoint's JWT token validation pipeline, letting unauthenticated attackers forge valid JWTs and impersonate any SharePoint site user or administrator. Threat actors began exploiting the bug after Rapid7 released a Python PoC that forges tokens and enumerates domain users by SID. KEVIntel telemetry shows 12 exploitation attempts since July 19, 2026, eight of them on August 12-13 following the PoC release, originating from eight IPs across Hong Kong, Japan, the Netherlands, Taiwan, and the U.S. This is the fifth SharePoint vulnerability exploited this year, alongside CVE-2026-45659, CVE-2026-56164, CVE-2026-58644, and CVE-2026-50522.

The Hacker News·Aug 15, 00:00 UTC · Aug 15, 2026Exploit / PoC in the wildCVE-2026-55040CVE-2026-45659CVE-2026-56164+2 CVEs

Quantifying the Engagement Trap: Impact of Short-form Video Recommender Systems on Users with ADHD

A 302-participant study finds engagement-optimized short-form video recommenders cause disproportionate time blindness and distress for users with ADHD.

Researchers ran a stratified Prolific study with 302 participants comparing short-form video recommendation experiences with and without ADHD. Participants with ADHD reported significantly higher time blindness, post-usage regret, and emotional distress despite perceiving recommendations as similarly relevant. The paper proposes neurodiversity-aware, human-centered design interventions to mitigate these algorithmic harms.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:38 UTC · 9d agoAI research

SAP Commerce Cloud CVE-2026-58231 Targeted in Exploitation Attempts Days After Patch

A maximum-severity (CVSS 10.0) unauthenticated RCE flaw in SAP Commerce Cloud, CVE-2026-58231, is under active exploitation days after patching.

CVE-2026-58231, rated 10.0 on CVSS, stems from insufficient authorization checks and input validation in SAP Commerce Cloud, enabling unauthenticated arbitrary code execution. Defused Cyber's honeypots detected exploitation attempts three days after the patch, and KEVIntel independently confirmed two attempts on August 14 from a US IP address. Onapsis urges customers to patch and rebuild, with an IP Filter Set offered as a temporary workaround. No actor attribution yet, though prior SAP NetWeaver flaw CVE-2025-31324 was exploited by China-nexus and criminal groups.

The Hacker News·Aug 17, 17:57 UTC · Aug 17, 2026Exploit / PoC in the wildCVE-2026-58231CVE-2025-31324

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection

Plan injection plants benign-sounding harmful reasoning that steers LLM actors to adversarial actions while evading chain-of-thought monitors.

Researchers show that injecting harmful but benign-sounding plans into an actor model's context causes it to perform adversarial actions while its reasoning passes chain-of-thought monitors, achieving 25-33% monitor evasion rates across benchmarks and scaling to larger models like DeepSeek-R1. Actor models paraphrase injected plans as their own reasoning without attribution. Giving the monitor access to the injected plan dropped detection by up to 50% on the Bio-Math task, with extra thinking tokens spent rationalizing rather than flagging the plan.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 18:33 UTC · 2d agoAI safety & security

SAP warns of maximum severity 'OVERPASS' kernel vulnerability

SAP's September 2026 patches fix maximum-severity kernel flaw CVE-2026-44756 (OVERPASS), letting unauthenticated attackers execute commands on 10,000+ internet-facing SAP systems.

SAP's September 2026 Patch Tuesday addresses 20 vulnerabilities, headlined by CVE-2026-44756, a buffer overflow in the Extended Passport Protocol (EPP) library of the SAP Kernel dubbed OVERPASS by Onapsis researchers. Unauthenticated attackers can gain administrative command execution via SAP Internet Communication Manager; Onapsis counted more than 10,000 internet-facing SAP systems. SAP also fixed CVE-2026-58240 (S4GET), a critical missing-authentication flaw in the NetWeaver Message Server enabling unauthenticated remote code execution as <sid>adm. Since November 2021, CISA has added 14 SAP flaws to its actively exploited list, three abused by ransomware gangs.

BleepingComputer·Sep 8, 14:55 UTC · 8d agoVulnerability 2 sourcesCVE-2026-44756CVE-2026-58240CVE-2026-58231

I accidentally turned LLM memory into program analysis

A pwning.systems write-up describes how LLM memory functionality was unexpectedly repurposed into a program analysis technique.

A security research post on pwning.systems describes the author's discovery that LLM memory behavior effectively functioned as program analysis. The write-up is hosted on a security-focused blog and surfaced via a security-tagged link aggregator. Detailed technical content is not included in this feed, limiting verifiable specifics.

Lobsters · security·Aug 29, 03:16 UTC · 18d agoResearch1

CISA Adds Six Exploited Flaws to KEV, Including NetScaler, Linux, and SQL Server Bugs

CISA added six actively exploited flaws to KEV, led by Citrix NetScaler CVE-2026-8452, with federal patch deadlines set.

CISA added six flaws to the KEV catalog on August 27, 2026, citing evidence of active exploitation, led by CVE-2026-8452, an improper buffer restriction in Citrix NetScaler ADC and Gateway that can cause denial-of-service. Defused Cyber and Previdian reported active exploitation of CVE-2026-8452, with 36 attempts over 12 days from 12 attacker IPs deploying web shells x.php and z.php. The other additions cover SQL Server RCE (CVE-2019-1068), Linux kernel out-of-bounds write (CVE-2022-0995), Red Hat ABRT and libuser privilege escalations (CVE-2015-5287, CVE-2015-3246), and AjaxPro deserialization RCE (CVE-2021-23758). FCEB agencies must patch CVE-2019-1068 and CVE-2026-8452 by August 29, 2026, and the rest by September 9, 2026.

The Hacker News·Aug 27, 07:05 UTC · 20d agoExploit / PoC in the wildCVE-2019-1068CVE-2026-8452CVE-2022-0995+3 CVEs

What Else Needs Fixing? Exploring Cost-Effective Test-Time Compute for Revision Propagation in Artifacts Generated Through Conversation

A new benchmark shows LLMs reach 68.3-93% accuracy propagating local revisions across conversationally generated artifacts, with parallel-sample selection most cost-effective.

The paper introduces a benchmark for revision propagation: when users request a local change, LLMs must identify dependencies and update all affected parts of an artifact generated through conversation, where context lives in the chat history. Nine revision methods, including sequential reflection and parallel sampling variants, were evaluated on gpt-oss-20b/120b, gpt-5.4-mini, and qwen3.5-9b/27b/122b. Baselines scored 68.3-93% accuracy, and selecting among three parallel samples via LLM-based or medoid selection improved accuracy by 2.2-9.7% as the most cost-effective test-time compute strategy. Code and dataset are released.

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

Cl0p Affiliates Target Internet-Exposed PTC Windchill and FlexPLM with Unauthenticated RCE

The Hacker News·Aug 3, 16:06 UTC · Aug 3, 2026Ransomware in the wildCVE-2026-12569

How Does mHC Use Its Residual Streams? Selective Routing and Near-Identity Mixing

Analysis of DeepSeek-V4-Flash shows four-stream mHC residual blocks use only about two streams effectively, with late-layer mixing providing little benefit.

The study examines the four-stream residual pathway of DeepSeek-V4-Flash, finding typical attention or FFN sites effectively use about two streams and that residual mixing is modest, occurring primarily in early layers. Replacing late mixers with identity increases C4 perplexity by only 1.9% while replacing early mixers raises it by 41%. Retaining the three largest routing weights per token increases perplexity by at most 2.7%, showing the model uses only part of the flexibility afforded by the four-stream design.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 16:02 UTC · 12d agoAI research

Attackers exploit critical SharePoint flaw after PoC goes public (CVE-2026-55040)

Attackers use Rapid7's public PoC to exploit CVE-2026-55040, a critical SharePoint authentication bypass patched in July 2026.

Rapid7 researcher Stephen Fewer published technical analysis and PoC exploit code for CVE-2026-55040, an authentication bypass in SharePoint caused by flaws in the JWT token validation pipeline. Threat intelligence firm Defused observed attackers using the PoC against SharePoint honeypots, though Microsoft has not yet confirmed in-the-wild exploitation. Chained with CVE-2026-63520, the flaw can yield unauthenticated remote code execution, per NHS England Digital. CISA is urging administrators to harden SharePoint deployments by limiting direct internet exposure and using Layer 7 reverse proxies.

Help Net Security·Aug 13, 00:00 UTC · Aug 13, 2026Exploit / PoC in the wildCVE-2026-55040CVE-2026-63520

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

Attackers Exploit VMware vCenter Vulnerability to Gain Persistent Remote Access

Attackers actively exploit critical VMware vCenter flaw CVE-2026-59310, installing reverse_ssh backdoors on 361 hosts across 47 countries.

QUIRSO observed active exploitation of CVE-2026-59310 (CVSS 9.8), a directory-traversal flaw in VMware vCenter allowing arbitrary code execution, patched by Broadcom in late July. Attack chains deploy a malicious cron job running the open-source reverse_ssh tool for persistence, with compromised systems phoning home starting August 3, five days after disclosure. Forensics show 361 unique victim IPs in 47 countries, mostly Germany, the U.S., Turkey, Iran, and France, with the actor suspected to be an APT. Separately, Defused Cyber reports rising scanning indicative of exploitation of related CVE-2026-59309, an unauthenticated auth bypass in vmdir.

The Hacker News·Aug 12, 16:41 UTC · Aug 12, 2026Exploit / PoC in the wildCVE-2026-59310CVE-2026-59309

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.CR·Sep 8, 07:38 UTC · 8d agoAI safety & security

Clop-Linked Windchill Web Shell Decrypts Credentials and Maps Engineering Data

ReliaQuest details a bespoke JSP web shell that Clop deploys on hacked PTC Windchill and FlexPLM servers after exploiting CVE-2026-12569.

ReliaQuest analyzed a custom Java web shell planted on vulnerable PTC Windchill and FlexPLM servers following exploitation of CVE-2026-12569 (CVSS 9.3). The implant decrypts Windchill keystore credentials including the LDAP manager password, enumerates the file vault for engineering data, and loads attacker-supplied Java classes in memory for post-exploitation. Commands let operators read and delete files, exfiltrate results, and deliver follow-on payloads such as ransomware. Ransom-ISAC, eCrime.ch and Defused previously attributed the campaign to the Clop data-theft extortion group.

The Hacker News·Aug 19, 15:27 UTC · 28d agoThreat actor in the wildCVE-2026-12569

Decomposition Buys Integrity, Not Yield

Study of 600 production deep-research traces finds agent-tree decomposition loses findings at rate N^(1-δ); flat architectures maximize yield.

The paper models multi-agent decomposition as a tree where an agent holding b items retains each with probability r(b); with r(b)=1/b every tree delivers exactly one finding regardless of shape. Analysis of 600 production deep-research traces estimates delta=0.34 retention decay, and 1,012 annotated traces show one brief in sixteen goes off-target per tier, giving an alignment penalty of 0.536. Depth still cuts root context exposure from N to N^(1/k) and is cheaper at scale, with a hazard model over 743,819 production tool calls showing delegation is an opening move rather than a response to filling context.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:04 UTC · 1d agoAI research

U.S. CISA adds Apple macOS, Microsoft SharePoint, Broadcom VMware vCenter, and Microsoft IKE flaws to its Known Exploited Vulnerabilities catalog

CISA adds actively exploited flaws in Microsoft IKE, SharePoint, VMware vCenter, and macOS Screen Sharing to KEV catalog, due August 21.

CISA added four vulnerabilities to its Known Exploited Vulnerabilities catalog: CVE-2026-33824 (Windows IKE remote code execution, CVSS 9.8), CVE-2026-55040 (SharePoint JWT authentication bypass, CVSS 9.1), CVE-2026-59310 (VMware vCenter Syslog path traversal, CVSS 9.8), and CVE-2026-65400 (macOS Screen Sharing improper authentication). The SharePoint flaw combines four weaknesses including alg:none JWT forging to impersonate any user, with Defused researchers observing attackers using Rapid7's PoC against honeypots. The Dutch NCSC confirmed active exploitation of the macOS Screen Sharing flaw on August 15, 2026. Federal agencies must patch by August 21, 2026.

Security Affairs·Aug 19, 07:19 UTC · 28d agoExploit / PoC in the wildCVE-2026-33824CVE-2026-55040CVE-2026-59310+1 CVEs

Inoculation Midtraining with Learned Neologisms

Inoculation Midtraining confines unsafe LLM behavior to a neologism-marked context, reducing misalignment after unsafe post-training but leaking under nearby contextual cues.

The paper introduces Inoculation Midtraining, which teaches a base model during midtraining that unsafe behavior belongs to a context marked by a learned neologism token, then post-trains on unsafe data within that context. Across supervised fine-tuning and RL post-training regimes, the technique reduces misalignment while preserving transfer of benign properties like German or Shakespearean prose. However, it does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. The authors conclude it is not yet a load-bearing component of a developer safety framework.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:12 UTC · 2d agoAI safety & security

⚡ Weekly Recap: VMware Exploits, Windows 0-Day, MCP Attacks, Browser Hijacks and More

Weekly recap: suspected China-nexus APT exploited VMware vCenter CVE-2026-59310, Lazarus used Windows zero-day CVE-2026-68820, and macOS flaw dropped Monero miners.

The week's top stories included a suspected China-nexus APT exploiting VMware vCenter CVE-2026-59310 (CVSS 9.8), a directory-traversal flaw enabling arbitrary code execution, with backdoors, reverse SSH and Babuk-derived ransomware deployed, per QUIRSO. NCSC-NL reported active exploitation of macOS Screen Sharing flaw CVE-2026-65400 (CVSS 9.8) to install Monero miners on systems with port 5900 exposed. Lazarus Group exploited Windows zero-day CVE-2026-68820 (CVSS 7.0) in Operation Dream Job, delivering ForestTiger and Troy backdoors to defense and aerospace targets. GeoServer patched a critical SQL injection flaw exploited within hours of disclosure, and Jamf analyzed the new Amnesia Stealer for macOS with live browser control.

The Hacker News·Aug 17, 17:57 UTC · Aug 17, 2026Exploit / PoC in the wildCVE-2026-59310CVE-2026-65400CVE-2026-688201

Refuse without Refusal: A Structural Analysis of Safety-Tuning Responses for Reducing False Refusals in Language Models

Study shows training LLMs on refusal rationales instead of boilerplate refusal statements reduces false refusals while maintaining safety performance.

The paper decomposes safety-tuning responses into a boilerplate refusal statement and an explanatory rationale, finding that refusal statements push models to rely on superficial cues and misjudge benign queries as harmful. Training solely on rationales reduces false refusals while maintaining comparable safety performance, and the benefits carry over to in-context learning configurations and remain compatible with inference-time mitigations. The results argue for precisely curated, fine-grained safety supervision datasets when aligning LLMs.

Hugging Face daily papers·Sep 3, 20:00 UTC · 13d agoAI safety & security1

Attackers plant remote access tools on compromised PaperCut servers

Attackers chained two PaperCut NG/MF zero-days for unauthenticated access, installing SimpleHelp and AnyDesk remote access tools on compromised servers.

An ongoing campaign exploits chained zero-days CVE-2026-81578 (improper access control) and CVE-2026-82078 (unsafe dynamic class loading) in internet-facing PaperCut NG and MF Application Servers, enabling authentication bypass and arbitrary Java bytecode execution. Post-compromise activity includes user and domain enumeration, payload download from sendit.sh, and silent installation of SimpleHelp and AnyDesk for redundant remote access; Defused observed CVE-2026-81578/CVE-2026-82078 exploit activity in honeypots since August 29, including data theft via Derby database dumps. Emergency patches were released August 28 and August 30, but 47% of roughly 2,500 PaperCut installs tracked by Huntress run v23 or older, for which no patch is available.

Help Net Security·Sep 2, 08:12 UTC · 14d agoExploit / PoC in the wildCVE-2026-81578CVE-2026-820781

VDiff-Bench: A Challenging Benchmark for Fine-Grained Image Difference Identification

VDiff-Bench, a 1,756-question benchmark, shows multimodal LLMs struggle with fine-grained image-difference identification, scoring as low as 8.7% on low-level changes.

VDiff-Bench is a multiple-choice benchmark of 1,756 four-way questions over image pairs covering 10 change categories including position, motion, color, texture, OCR/text and illumination, with curated hard negatives. Evaluation of 11 state-of-the-art open- and closed-source MLLMs shows fine-grained visual comparison remains brittle: 7-8B-scale open-source models score 52.5-70.6% on semantic changes but only 8.7-33.3% on low-level changes like noise and texture. Notably, Grok 4.3 shows a sharp performance drop on noise and texture differences, falling behind large open-source models like Kimi K2.5 and K3.

Hugging Face daily papers·Sep 4, 20:00 UTC · 12d agoAI research

Previously patched Citrix NetScaler flaw exploited in the wild (CVE-2026-8452)

Actively exploited Citrix NetScaler flaw CVE-2026-8452, patched in June, now added to CISA KEV; attackers deploy web shells and run discovery commands.

CISA added CVE-2026-8452, a memory overflow vulnerability in Citrix NetScaler ADC/Gateway, to its Known Exploited Vulnerabilities catalog on August 26, 2026, giving federal agencies until August 29 to remediate. Citrix disclosed and patched the flaw on June 30, 2026 in versions 14.1-72.61, 13.1-63.18, and 13.1-37.272. watchTowr Labs found the flaw can be chained into unauthenticated remote code execution and released a PoC on August 14. Previdian and Defused reported exploitation with web shells named x.php and z.php from at least three unique source IPs.

Help Net Security·Aug 27, 00:00 UTC · 20d agoExploit / PoC in the wildCVE-2026-8452CVE-2015-3246CVE-2015-5287+3 CVEs

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.CL·Sep 15, 16:43 UTC · 1d agoAI research

Can We Stop The Ads? Taxonomy and Characterization of Smartphone Splash Ads and Existing Countermeasures

Study of 108 ad-defense implementations finds only one tool blocked splash-ad navigation across ten popular apps, and it required Accessibility permission.

The paper taxonomizes smartphone splash ads — full-screen ads at app launch that trick users into trigger mechanisms such as moving the phone — and analyzes 108 documented advertising defenses for deployment barriers. Many defenses require device rooting, jailbreaking, runtime code injection, or application modification; others need extra permissions, rule maintenance, compilation, or payment. In evaluating 13 configurations of 11 tools across 10 popular apps, only one prevented ad-triggered navigation across all ten apps, requiring Accessibility permission and leaving ads visible roughly one second before dismissal. Documented harms include delayed emergency response, driver distraction, and degraded accessibility for vision-impaired users.

arXiv cs.CR·Sep 15, 15:24 UTC · 1d agoResearch

RelateAnything: Real-Time Open-Vocabulary Relation Prediction From Any Inputs

RelateAnything is a 53M-parameter open-vocabulary relation prediction model running at 20 ms/frame, with 2.3-3.5x higher mean recall than comparable open-vocabulary methods.

RelateAnything predicts scored relations between image regions using any predicate vocabulary supplied at inference as text embeddings, with object labels never required as input, so region sources can change without retraining. Training covers 19,103 predicates using positive-unlabeled supervision; the authors release RA-4M (474k images, 4.3M geometrically verified relations over 10,102 free-text predicates) and the OV-SGG-Bench evaluation suite. The 53M-parameter model runs at 20 ms/frame and achieves 2.3-3.5x the mean recall of the strongest comparable open-vocabulary method across cross-dataset and zero-shot benchmarks. Model, corpus, and benchmark are public.

Hugging Face daily papers·Sep 10, 20:00 UTC · 6d agoAI research