Researchers expose transcript consistency flaws in major E2EE group

🕵️ RESEARCH & DEEP DIVES

• Researchers expose transcript consistency flaws in major E2EE group chats
  Major end-to-end encrypted group chat apps lack transcript consistency, enabling selective message manipulation.

  • Applies to major E2EE messaging apps with group chats, including Signal, iMessage, WhatsApp, and Threema
  • Vulnerability allows malicious group members to omit, reorder, or alter messages selectively per recipient
  • Exploitation enables social engineering, moderation evasion, and rigging of group polls
  • Attack leverages protocol fallback paths and pairwise delivery channels within groups
  Sources 📎 Coverage: arxiv.org · 📄 Original: arxiv.org · 👁 via arXiv cs.CR

• New HMM-Based Method Analyzes Multi-Phase Cyber Attacks on IEC 61850 Digital Substations
  Researchers developed a model to infer phases of cyber attacks on digital substations using IDS data.

  • Applies to digital substations compliant with IEC 61850 in modern power systems
  • Targets manipulation of circuit breaker operations via IEC 61850 communication
  • Attacks use multi-phase strategies that existing IDS detect only as isolated symptoms
  • Proposed SubCASP method uses Hidden Markov Model to fuse IDS logs for attack phase inference
  • Model trained on attack-graph datasets and tested under varying IDS observability and missing data
  Sources 📎 Coverage: arxiv.org · 📄 Original: arxiv.org · 👁 via arXiv cs.CR

• GradLock attack injects private data into AI models via compromised open-source components
  Researchers reveal GradLock, a training-time attack that embeds private data into AI model parameters.

  • Applies to AI models trained using compromised open-source components in the supply chain
  • GradLock injects sensitive training data directly into model parameters during training
  • Uses stateless deterministic indexing and dynamic gradient locking to preserve injected data
  • Enables near-lossless extraction of private data from final models without training environment access
  • Robust against quantization, pruning, and fine-tuning; evades detection by 93.3% of users
  Sources 📎 Coverage: arxiv.org · 📄 Original: arxiv.org · 👁 via arXiv cs.CR

• clean-label-dormant-to-activated-backdoor-machine-unlearning
  📎 Original: arxiv.org

🔓 CVEs & KEV

• CVE-2026-18157 — CVSS 7.8 — Yggdrasil-worker-package-manager: remote code execution vulnerability
• CVE-2026-6890 — Use of default credentials vulnerability
• CVE-2026-6889 — Denial of service vulnerability in Advan...
• CVE-2026-14541 — Authentication Bypass and Audience Confusion in MCP Toolbox OAuth Provider
• CVE-2026-14540 — Server-Side Request Forgery via Unrestricted HTTP Redirection in MCP Toolbox
• CVE-2026-14539 — Denial of Service via Unrestricted Payload Buffering in MCP Toolbox
• CVE-2026-14538 — BigQuery Dataset Allowlist Bypass via Metadata Dry-Run in MCP Toolbox
• CVE-2026-14537 — Authorization Bypass in MCP Toolbox Legacy HTTP Endpoints