All news with #research tag

⚠️ Recent research demonstrates that small amounts of narrow finetuning can produce broad, unexpected shifts in LLM behavior. The authors show weird generalization—models adopting outdated worldviews from bird-naming examples—and introduce inductive backdoors, where models learn triggers and behaviors via generalization. These effects enable persona hijacking and hard-to-detect misalignment.

⚠️ AI-powered VS Code forks such as Cursor, Windsurf, Google Antigravity and Trae were found recommending extensions that do not exist in the Open VSX registry, creating unclaimed namespaces attackers could register. Koi researcher Oren Yomtov showed that a single click on a suggested install (for example, a placeholder ms-ossdata.vscode-postgresql) can deploy a rogue package, and one placeholder received over 500 installs. Cursor and Google have released fixes, and the Eclipse Foundation removed non-official contributors and tightened registry safeguards. Developers should verify publishers before accepting IDE extension recommendations.

🔍 Cloudflare analyzed a BGP route leak observed on January 2 involving AS8048 (CANTV) redistributing prefixes originated by AS21980 (Dayco Telecom) via upstreams including AS6762 (Sparkle) and AS52320 (V.tal/GlobeNet). The pattern — with eleven similar events since December, heavy AS prepending, and an upstream provider relationship — suggests misconfigured export/import policies rather than deliberate interception. ROV would not have prevented this path-based leak; adoption of ASPA, RFC9234/OTC, and Peerlock-style checks is recommended to mitigate future leaks.

🛡️ In a recent interview Trend Micro COO Kevin Simzer described how a digital twin — a virtual replica built from enterprise telemetry — lets organizations run safe, comprehensive red-team simulations across real-world topologies. The approach enables what-if analyses, testing of security controls and architectural changes without risk to production systems. Simzer also noted additions like agentic capabilities to automate SIEM integration and Trend's plan to train proprietary AI models from its historical threat data.

🛡️ CrowdStrike outlines its methodology for training security-focused GenAI models at scale using the Google Cloud Vertex Training Cluster and an infrastructure-as-code approach. The team leverages Slurm for workload scheduling, modular data pipelines with synthetic augmentation, and a mix of parallelism strategies (data, tensor, pipeline, sequence/expert) to match model size and hardware. They optimize across GPU architectures (H100, B200) using high-performance attention kernels like Flash Attention and NCCL for inter-node communication to improve throughput, support extended contexts, and manage memory via gradient checkpointing and observability tooling.

📘 The Future Report, produced with youth consultancy Livity, surveyed over 7,000 teenagers (13–18) across France, Greece, Ireland, Italy, Poland, Spain and Sweden about their digital lives and expectations. It finds that 40% use AI daily or almost daily and that 81% of users report AI improved aspects of learning or creativity. Teens are largely optimistic yet express concerns about over-reliance, skill erosion and information trustworthiness. The report recommends stronger digital literacy, safety measures and meaningful youth participation in design and policy.

🔓Researchers revealed new XML-parsing exploits that severely weaken SAML-based SSO, demonstrating full authentication bypass against popular Ruby and PHP SAML libraries. PortSwigger researcher Zak Fedotkin presented these techniques at Black Hat Europe and published an open-source toolkit to identify and reproduce affected deployments. The work highlights attack vectors such as attribute pollution, namespace confusion, and a new class of void canonicalization that can circumvent XML signature validation. While fixes (including updates to Ruby-SAML) have been released, Fedotkin warns that only a foundational rework of SAML libraries will eliminate these systemic weaknesses.

🔓 Researchers at KU Leuven built a $50 DDR4 interposer that subverts confidential computing protections such as Intel SGX and AMD SEV, demonstrated at Black Hat Europe. The runtime attack, called Battering RAM, manipulates memory address mapping to gain arbitrary plaintext read/write and extract SGX provisioning keys, circumventing recent boot-time mitigations. The team warns that compromised memory modules in the supply chain could enable persistent backdoors on vulnerable cloud VMs.

🔍 IDC warns of a growing Total Cost of Ownership (TCO) crisis as AI inference becomes the dominant workload. Their global survey of 1,300 AI decision-makers finds inference already accounts for 47% of AI operations and is magnified by agentic workflows that trigger many sequential model calls. The research attributes the problem to fragmented stacks and idle accelerators and recommends shifting to integrated, system-level architectures that unite software, storage, networking, and compute. Google Cloud highlights AI Hypercomputer as a purpose-built solution to improve utilization and cost-effectiveness.

🔍 The Cybersecurity and Infrastructure Security Agency (CISA), with MITRE/HSSEDI, released the 2025 CWE Top 25, highlighting the most exploited software weaknesses that enable data theft, system compromise, and service disruption. The list is designed to help developers, security teams, and procurement managers prioritize fixes and adopt Secure by Design practices. CISA urges organizations to integrate the Top 25 into vulnerability management and procurement decisions to reduce risk and downstream costs.

🤖 The 2025 DORA companion guide highlights that AI acts as an amplifier, boosting strengths and exposing weaknesses across teams. Drawing on a cluster analysis of nearly 5,000 technology professionals, it identifies seven foundational capabilities — including a clear AI stance, healthy and AI-accessible data, strong version control, small-batch workflows, user-centric focus, and quality internal platforms — that increase the odds of positive outcomes. The guide maps seven team archetypes to help leaders diagnose where to start and offers a Value Stream Mapping facilitation to direct efforts toward system-level constraints so AI-driven productivity scales safely.

🔁 The article traces redundancy from tangible rack-level practices to fragile cloud and software-defined environments. It argues that physical diversity, disciplined configuration management and automation remain essential as networks span BGP, SD-WAN, edge devices and cloud control planes. Real resilience requires policy alignment, diverse DNS and routing protections and rehearsed pre-mortems so backups are usable when they matter most.

⚠️ Gartner has advised enterprises to block AI browsers until associated risks can be adequately managed. In its report Cybersecurity Must Block AI Browsers for Now, analysts warn that default settings prioritise user experience over security and list threats such as prompt injection, credential exposure and erroneous agent actions. Researchers and vendors have also flagged vulnerabilities and urged risk assessments and oversight.

🛡️ Chrome describes a layered approach to secure agentic browsing with Gemini, focusing on defenses against indirect prompt injection and goal‑hijacking. A new User Alignment Critic — an isolated, high‑trust model — reviews planned agent actions using only metadata and can veto misaligned steps. Chrome also enforces Agent Origin Sets to limit readable and writable origins, adds deterministic confirmations for sensitive actions, runs prompt‑injection detection in real time, and sustains continuous red‑teaming and monitoring to reduce exfiltration and unwanted transactions.

🧩 Bruce Schneier highlights a new paper proposing the Naibbe cipher, a verbose homophonic substitution method that transforms Latin and Italian plaintext into ciphertext resembling the Voynich Manuscript. The author demonstrates the cipher can be executed entirely by hand with plausible 15th‑century materials. Applied to a range of texts, Naibbe reproduces many of the manuscript’s key statistical properties while remaining decipherable. Schneier observes this keeps the ciphertext hypothesis viable and places constraints on plausible substitution structures.

🔗 The article argues that traditional threat feeds no longer suffice in modern, interconnected environments and proposes a Unified Linkage Model (ULM) to transform static indicators into dynamic threat flows. ULM defines three core linkage types — adjacency, inheritance and trustworthiness — to map how risk propagates across systems. It outlines practical steps to ingest and normalize feeds, establish and score linkages, integrate with MITRE ATT&CK and risk frameworks, and visualize attack pathways for prioritized response and compliance.

🔎 The UK’s National Cyber Security Centre and Canada’s Centre for Cyber Security (CCCS) have published a report on public content provenance aimed at improving digital trust in the AI era. It examines emerging provenance technologies, including trusted timestamps and cryptographically secured metadata, and identifies interoperability and usability gaps that hinder adoption. The guidance offers practical steps for organisations considering provenance solutions.

🔍 A newly discovered phishing framework named GhostFrame has been linked to more than one million attacks, according to Barracuda. The kit uses a benign-looking outer HTML page that conceals a malicious iframe, enabling attackers to swap content, target regions and evade scanners without changing the visible landing page. GhostFrame employs a two-stage chain: the loader creates randomized subdomains and validates them before loading an internal credential-stealing page, and includes anti-analysis controls that block inspection shortcuts and restrict user actions. Barracuda recommends a multilayered defense—regular browser updates, staff training, email gateways and web filters, restricting iframe embedding, and monitoring for injected or redirected content.

🛡️ CISA launched the Industry Engagement Platform (IEP) to create a structured, two-way channel between the agency and companies, researchers, and academia to present emerging cybersecurity and infrastructure technologies. The platform lets organizations build customizable technology profiles and upload capability overviews to connect with the right CISA subject-matter experts. Participation does not confer preferential contract consideration, but informs CISA market research and mission needs.

🔍 ISC2’s 2025 Cybersecurity Workforce Study, based on responses from more than 16,000 professionals, reports that 59% of organizations now face critical or significant cyber-skills shortages, up from 44% last year. Technical gaps are most acute in AI (41%), cloud security (36%), risk assessment (29%) and application security (28%), with governance, risk and compliance and security engineering each at 27%. The survey cites a dearth of talent (30%) and budget shortfalls (29%) as leading causes and links shortages to concrete impacts—88% reported at least one significant security incident. Despite concerns, headcount appears to be stabilizing and many professionals view AI as an opportunity for specialization and career growth.