All news with #research tag

🛡️ Recent research shows the “Whisper Leak” attack can infer the topic of LLM conversations by analyzing timing and packet patterns during streaming responses. Microsoft’s study tested 30 models and thousands of prompts, finding topic-detection accuracy from 71% to 100% for some models. Providers including OpenAI, Mistral, Microsoft Azure, and xAI have added invisible padding to network packets to disrupt these timing signals. Users can further protect sensitive chats by using local models, disabling streaming output, avoiding untrusted networks, or using a trusted VPN and up-to-date anti-spyware.

🔒 Submarine cables carry between 95% and 99% of global data traffic, yet recent breakages — notably ten in the Baltic Sea between 2022 and July 2025 — highlight persistent vulnerabilities. Private operators now control most capacity, and governments and vendors must address both physical threats such as fishing and anchors and increasingly sophisticated cyber risks. Major cloud vendors emphasize route diversity and redundancy while operators like Telxius combine burial, audits, AI/ML detection and continuity planning to protect service availability.

🔐 The article explains how indirect prompt injection — malicious instructions embedded in external content such as documents, images, emails and webpages — can manipulate AI tools without users seeing the exploit. It contrasts indirect attacks with direct prompt injection and cites CrowdStrike's analysis of over 300,000 adversarial prompts and 150 techniques. Recommended defenses include detection, input sanitization, allowlisting, privilege separation, monitoring and user education to shrink this expanding attack surface.

✍️ Researchers from Icaro Lab (DexAI), Sapienza University of Rome, and Sant’Anna School found that short poetic prompts can reliably subvert AI safety filters, in some cases achieving 100% success. Using 20 crafted poems and the MLCommons AILuminate benchmark across 25 proprietary and open models, they prompted systems to produce hazardous instructions — from weapons-grade plutonium to steps for deploying RATs. The team observed wide variance by vendor and model family, with some smaller models surprisingly more resistant. The study concludes that stylistic prompts exploit structural alignment weaknesses across providers.

🧬 Combining Google’s ADK, Gemini, and Cloud infrastructure, this work reframes variant interpretation as a conversational workflow that removes repetitive scripting and context switching. A two-phase design performs heavy VEP annotation once, stores versioned ADK artifacts and public BigQuery datasets, and enables sub-5-second interactive queries via a QueryAgent. Validation with an APOB spike-in demonstrated single-variant precision, compatibility across DeepVariant versions, and scalability to ~8.8M variants.

🛡️ Zero-day attacks exploit software vulnerabilities that are unknown to the vendor, enabling attackers to compromise systems before patches are available. They target high-value platforms such as operating systems, web browsers, enterprise applications, and IoT devices, often using spear-phishing or zero-click techniques. Because signature-based tools frequently miss novel exploits, effective defense requires rapid patching, behavior-based detection (EDR, NDR, XDR), network segmentation, and investigative analysis of packet-level data to detect, contain, and learn from incidents.

⚠️ Researchers report that converting prompts into poetry can reliably jailbreak large language models, producing high attack-success rates across 25 proprietary and open models. The study found poetic reframing yielded average jailbreak success of 62% for hand-crafted verses and about 43% for automated meta-prompt conversions, substantially outperforming prose baselines. Authors map attacks to MLCommons and EU CoP risk taxonomies and warn this stylistic vector can evade current safety mechanisms.

🔒 Singaporean researchers demonstrated how inexpensive offline dashcams can be weaponized into a self‑propagating surveillance network. They identified common weaknesses — default or hardcoded Wi‑Fi credentials, exposed services (FTP/RTSP), MAC‑spoofing and replay attacks — that allow attackers to download video, audio, timestamps and GPS metadata. The team showed mass compromise is feasible and offered mitigation steps for vendors and drivers.

🛡️ This edition of Talos Threat Source urges a simple behavioral shift: practice care in what, how, and why you share information during the holiday season and beyond. The briefing highlights operational pressures as teams run lean and attackers intensify phishing and supply‑chain campaigns, and it outlines practical changes such as retiring obsolete ClamAV signatures and encouraging feature‑release container tags for better security maintenance. Thoughtful, timely sharing of tips, IOCs, and status updates can materially improve collective resilience when resources are constrained.

🛡️ Researchers at Huntress uncovered a ClickFix campaign that hides malware inside the RGB pixels of PNG images on a fake Windows Update page, tricking victims into pasting and running commands. The delivered payloads include the LummaC2 infostealer and the Rhadamanthys malware family, with active domains observed after a mid-November takedown. Huntress warns the steganographic technique and the realistic Windows Update motif increase the attack's stealth, and recommends disabling the Windows Run dialog and strengthening endpoint monitoring.

🤖 Agentic AI is emerging as a practical accelerator for security teams, automating detection, triage, remediation and routine operations to improve speed and scale. Security leaders at Zoom, Dell, Palo Alto and others highlight its ability to reduce alert fatigue, augment SOCs and act as a force multiplier amid persistent skills shortages. Implementations emphasize augmentation over replacement, enabling continuous monitoring and faster, more consistent responses.

🔒 Security researchers at watchTowr discovered that two popular code utility sites — JSON Formatter and Code Beautify — inadvertently exposed thousands of developer submissions containing sensitive secrets and credentials. By querying a public API and the sites’ “Recent Links” listings, the team extracted over 80,000 submissions spanning years, including API keys, private keys, database and cloud credentials, JWTs, and PII. The exposure remained until the sites disabled the save feature; watchTowr also confirmed active scraping by third parties and reported limited response from affected organizations.

🔐 Tor has replaced its legacy tor1 relay encryption with a new design called Counter Galois Onion (CGO) to strengthen circuit traffic confidentiality and integrity. CGO is built on a Rugged Pseudorandom Permutation (RPRP) construction named UIV+ and provides wide-block encryption, tag chaining, per-cell key updates for immediate forward secrecy, and a 16-byte authenticator that removes SHA-1. The change is currently experimental in the C Tor implementation and the Rust client Arti, will be deployed transparently to Tor Browser users, and aims to block tagging and other malleability attacks with only modest bandwidth cost.

🛡️ Microsoft Defender Experts outlines how autonomous AI agents are transforming Security Operations Centers by automating repetitive triage and amplifying analyst impact. Built with expert-defined guardrails, curated test sets, and human-in-the-loop validation, these agents already process about 75% of phishing and malware cases and help resolve incidents nearly 72% faster. The program emphasizes human governance, auditability, and iterative rollout through dark-mode evaluation and pilot partnerships.

🔒 New research from watchTowr Labs found over 80,000 files saved to online code-formatting tools, exposing thousands of passwords, API keys, repository tokens and other sensitive credentials across government, telecoms, finance, healthcare and critical infrastructure. The datasets comprise five years of JSONFormatter content and one year of CodeBeautify content (about 5GB), and both services used predictable, shareable URLs and a Recent Links page that made mass crawling trivial. Researchers uploaded decoy AWS keys that were abused within 48 hours, and both sites have temporarily disabled save functionality while implementing enhanced content-prevention measures.

🔒 Telecom operators must abandon perimeter assumptions and adopt a zero trust mindset that treats verification as continuous rather than a one-time event. This shift is organizational as much as technical, requiring unified IT/OT policies, least-privilege access and microsegmentation to limit lateral movement. The article recommends pragmatic steps — wrapping legacy systems with secure gateways and centralized authentication — and aligning controls with frameworks such as NIST and NIS2, while tracking concrete KPIs in the first 180 days.

🚀 Google Cloud engineers demonstrated an experimental GKE cluster running 130,000 nodes to validate extreme scalability for AI/ML workloads. The test sustained control-plane throughput near 1,000 operations per second, supported over one million datastore objects, and achieved a baseline of 130,000 Pods launching in 3 minutes 40 seconds. The project combined API-server caching KEPs, a Spanner-backed key-value storage backend, and job-level orchestration via Kueue to enable predictable admission, rapid preemption, and efficient utilization at massive scale.

🔒 Security researchers have uncovered Operation WrtHug, a global campaign that has hijacked thousands of largely end-of-life ASUS WRT routers by chaining at least six known vulnerabilities. Over roughly six months analysts identified about 50,000 unique infected IPs, predominantly in Taiwan, using a distinctive malicious self-signed AiCloud certificate with a 100-year lifetime as an indicator of compromise. Owners are urged to apply ASUS firmware updates or replace unsupported models and disable remote-access features to mitigate risk.

🔒 Kendra Albert's USENIX talk, highlighted by Bruce Schneier, argues that modern managed bug bounty programs often impose contractual confidentiality that prevents researchers from publicly sharing vulnerabilities. These restrictions can flip the original bargain of coordinated vulnerability disclosure, silencing researchers while allowing vendors to delay or avoid fixes. Schneier urges platforms and companies to prohibit mandatory non‑disclosure terms and restore the balance between researcher reporting and vendor remediation.

🔒 The Immersive Cyber Workforce Benchmark Report 2025 warns that cyber readiness is stalling despite increased confidence in incident response: resilience scores have remained flat since 2023 and the median time to complete critical exercises is 17 days. In the Orchid Corp crisis scenario participants averaged 22% decision accuracy and took 29 hours to contain incidents. Immersive highlights that only 41% of organisations include non-technical roles in simulations and that 60% of training focuses on CVEs older than two years, urging regular, completed training, senior leadership involvement and a focus on current threats and the three pillars: prove, improve, report.