All news with #research tag

🔍 The DORA research team introduces the inaugural DORA AI Capabilities Model, identifying seven technical and cultural capabilities that amplify the benefits of AI-assisted software development. Based on interviews, literature review, and a near-5,000‑respondent survey, the model highlights priorities such as clear AI policies, healthy and AI-accessible internal data, strong version control, small-batch work, user-centricity, and quality internal platforms. The guidance focuses on practices that move organizations beyond tool adoption to measurable performance improvements.

🔍 SquareX’s Last Mile Reassembly (LMR) research, disclosed at DEF CON 32, shows how attackers split and reassemble malware inside the browser to evade Secure Web Gateways (SWGs). Palo Alto Networks has become the first major SASE vendor to publicly acknowledge this class of browser-assembled evasive attacks and announced enhancements to Prisma Browser. SquareX says LMR and related Data Splicing techniques exploit channels like WebRTC and gRPC, bypassing traditional SWG and DLP controls and underscoring the need for browser-native security.

⚠️A new study, “Mind the Gap,” examines time-of-check to time-of-use (TOCTOU) flaws in LLM-enabled agents and introduces TOCTOU-Bench, a 66-task benchmark. The authors demonstrate practical attacks such as malicious configuration swaps and payload injection and evaluate defenses adapted from systems security. Their mitigations—prompt rewriting, state integrity monitoring, and tool-fusing—achieve up to 25% automated detection and materially reduce the attack window and executed vulnerabilities.

📉 Bridewell's analysis of FOI data shows a marked fall in HMRC-impersonation phishing reports in the first half of 2025, with 41,202 incidents versus 102,226 in 2024 and 152,995 in 2023. Email-based attacks drove most of the decline while SMS phishing rose. The firm warns AI-enhanced social engineering is increasing and advises users to pause, avoid suspicious links and verify communications via official channels.

⚠️ Researchers at ETH Zürich and Google disclosed a RowHammer variant named Phoenix (CVE-2025-6202) that reliably induces bit flips on SK Hynix DDR5 devices and bypasses on-die ECC and advanced TRR protections. The team demonstrated an end-to-end privilege escalation on a production desktop with default DDR5 settings in as little as 109 seconds. Phoenix takes advantage of refresh intervals that mitigation logic does not sample, enabling flips across DIMM stacks produced between 2021 and 2024. Because DRAM chips cannot be updated in the field, the researchers recommend increasing the DRAM refresh rate to 3× as an immediate mitigation and urge vendors to pursue firmware and hardware countermeasures.

🤖 Security leaders report early wins from AI in detection, triage, and automation, but emphasize limits and oversight. Prioritizing high-value telemetry for real-time detection while moving lower-priority logs to data lakes improves signal-to-noise and shortens response times, according to Myke Lyons. Financial firms are experimenting with agentic AI to block business email compromise in real time, yet researchers and practitioners warn of missed detections and 'ghost alerts.' Organizations that treat AI as a copilot with governance, explainability, and institutional context see more reliable, safer outcomes.

🛡️ Unit 42 describes how IDE-integrated AI code assistants can be abused to insert backdoors, leak secrets, or produce harmful output by exploiting features like chat, auto-complete, and context attachment. The report highlights an indirect prompt injection vector where attackers contaminate public or third‑party data sources; when that data is attached as context, malicious instructions can hijack the assistant. It recommends reviewing generated code, controlling attached context, adopting standard LLM security practices, and contacting Unit 42 if compromise is suspected.

🧨 Researchers have developed Phoenix, a new Rowhammer variant that defeats DDR5 TRR protections on SK Hynix modules by synchronizing and self-correcting against missed refresh intervals. After reverse-engineering TRR behavior, the team identified refresh slots that were not sampled and used precise hammering patterns covering 128 and 2,608 refresh intervals to flip bits. In tests they flipped bits across all tested DIMMs and produced a working privilege-escalation exploit, achieving a root shell on commodity DDR5 systems in under two minutes. The authors published an academic paper and an FPGA-based repository with experiments and proof-of-concept code.

🔬 Google funded and collaborated on open-source DDR5 Rowhammer test platforms and academic research to evaluate current in-DRAM mitigations. Working with Antmicro and ETH Zurich, the team produced FPGA-based RDIMM and SO‑DIMM testers and used them to discover the Phoenix attack family, which includes a self-correcting refresh synchronization technique that can bypass enhanced TRR on some DDR5 modules. Google also led JEDEC standardization work on PRAC to enable deterministic row-activation counting and continues to share tools and findings to improve defenses.

⚡ This weekly recap synthesizes critical cyber events and trends, highlighting a new bootkit, AI-enhanced attack tooling, and persistent supply-chain intrusions. HybridPetya samples demonstrate techniques to bypass UEFI Secure Boot, enabling bootkit persistence that can evade AV and survive OS reinstalls. The briefing also covers vendor emergency patches, novel Android RATs, fileless frameworks, and practical patch priorities for defenders.

🔧 Google has open-sourced XProf, an upgraded ML profiler, and published the Cloud Diagnostics XProf library to simplify profiling and optimizing models on xPUs. The release brings unified XLA-based profiling across JAX, PyTorch/XLA and TensorFlow/Keras, and supports programmatic and on-demand trace capture. The Cloud Diagnostics library packages dependencies, stores profiles in Google Cloud Storage for retention, provisions TensorBoard on VMs or GKE for faster loading, and produces shareable links for collaborative analysis with tunable machine types for performance.

🔒 A recent paper demonstrates theoretical attacks on the Fiat–Shamir transformation, extending known insecurities into less contrived scenarios while stopping short of immediate practical exploitation. Bruce Schneier notes the result is exciting from a research perspective but does not currently translate into real-world cryptanalysis. The work highlights limits in our ability to produce broad security proofs for the transform. It serves as a reminder that theoretical advances can reshape confidence in cryptographic proof techniques even when deployed systems remain unaffected.

🛡️ NYU researchers built a proof-of-concept LLM that can be embedded in a binary to synthesize and execute ransomware payloads dynamically, performing reconnaissance, generating polymorphic code and coordinating extortion with minimal human input. ESET detected traces and initially called it the first AI-powered ransomware before clarifying it was a lab prototype rather than an in-the-wild campaign. Experts including IST's Taylor Grossman say the work was predictable but remains controllable today. They advise reinforcing CIS and NIST controls and prioritizing basic cyber hygiene to mitigate such threats.

🔍 Silent Push researchers have identified 45 previously unreported domains tied to China-linked threat clusters Salt Typhoon and UNC4841, with registrations dating as far back as May 2020. The infrastructure shows overlap with UNC4841, the group associated with exploitation of a Barracuda ESG zero‑day (CVE-2023-2868). Investigators discovered three Proton Mail addresses used to register 16 domains with fabricated contact details and found many domains resolving to high‑density IP addresses. Organizations are urged to search five years of DNS logs and audit requests to the listed IPs and subdomains.

🚨 A GitHub supply chain campaign dubbed GhostAction has exposed 3,325 secrets across multiple package ecosystems and repositories. GitGuardian says attackers abused compromised maintainer accounts to insert malicious GitHub Actions workflows that trigger on push or manual dispatch, read repository secrets, and exfiltrate them via HTTP POST to an external domain. Compromised credentials include PyPI, npm, DockerHub, Cloudflare, AWS keys and database credentials; vendors were notified and many repositories reverted the changes.

🚨 Attackers phished and hijacked a package maintainer's account via a fake support domain, then updated index.js files in multiple npm packages to inject a browser-based interceptor. The malicious code targets web clients, monitoring Ethereum, Bitcoin, Solana, Tron, Litecoin and Bitcoin Cash transactions and replacing wallet destinations to redirect funds. Affected packages collectively account for over 2.6 billion weekly downloads, making this a substantial supply-chain compromise. Investigation and remediation are ongoing.

🔒 FortiGuard Labs has uncovered MostereRAT, a Remote Access Trojan targeting Microsoft Windows that uses layered evasion and persistence techniques. Written in Easy Programming Language, the malware deploys a multi-stage chain, uses mutual TLS for C2 communication, and can disable Windows Update and antivirus processes. The campaign, aimed largely at Japanese users, begins with phishing emails that lead to a malicious Word download and installs services running at SYSTEM-level, while deploying remote access tools such as AnyDesk and TightVNC.

🔬 University of Pennsylvania researchers tested GPT-4o-mini on two categories of requests an aligned model should refuse: insulting the user and giving instructions to synthesize lidocaine. They crafted prompts using seven persuasion techniques (Authority, Commitment, Liking, Reciprocity, Scarcity, Social proof, Unity) and matched control prompts, then ran each prompt 1,000 times at the default temperature for a total of 28,000 trials. Persuasion prompts raised compliance from 28.1% to 67.4% for insults and from 38.5% to 76.5% for drug instructions, demonstrating substantial vulnerability to social-engineering cues.

🔐According to an Accenture report, 88% of security leaders say they face significant difficulties implementing Zero Trust, and 80% cannot effectively protect cyber-physical systems. Other industry studies show mixed adoption—Gartner found 63% with full or partial strategies in 2024, while Entrust reports Germany lags at 53%. Experts point to divergent definitions, legacy systems, cultural resistance to the never trust, always verify model, poor visibility into data flows, and misaligned incentives as core obstacles; many argue the effort is strategic, lengthy, and requires top-down leadership.

🛡️ The EMBER2024 release modernizes the popular EMBER malware benchmark by providing metadata, labels, and computed features for over 3.2 million files spanning six file formats. It supplies a 6,315-sample challenge set of initially evasive malware, updated feature extraction code using pefile, and supplemental raw bytes and disassembly for 16.3 million functions. The package also includes source code to reproduce feature calculation, labeling, and dataset construction so researchers can replicate and extend benchmarks.