Belgian eID software flaws exposed 2 million users to data theft

🕵️ RESEARCH & DEEP DIVES

• Belgian eID software flaws exposed 2 million users to data theft and RCE Belgian eID software flaws exposed users to identity theft and remote code execution.

  • Connective served over 2 million Belgian users, including customers of 8 of the country's 10 largest banks and 60+ government agencies.
  • The browser extension and native host could expose eID and Maestro card data and enable forged legally binding signatures.
  • Missing origin authentication let any website or embedded ad communicate with the local Connective application.
  • Attacker-controlled PIN prompts exposed eID PINs, while a drive-by file attack could execute attacker-supplied code at user level without an eID card connected. 📄 Source: amibeingpwned.com · 📎 Coverage: securityweek.com · 👁 via SecurityWeek

• Fairis targets fairness-poisoning attacks in collaborative machine learning Researchers introduced Fairis to contain fairness-poisoning attacks in collaborative machine learning.

  • The research targets collaborative machine learning among financial institutions.
  • A malicious client can maximize group disparity while preserving model accuracy.
  • The attacker manipulates local updates and can game FairFed's gap-based weighting by observing the global fairness score.
  • Fairis applies server-side reweighting based on each client's local Equal Opportunity Difference.
  • In Taiwan Credit tests, Fairis reduced a stealthy adversary's weight by 41% to 54% below a size-blind control. 📄 Source: csrc.nist.gov · 📎 Coverage: arxiv.org · 👁 via arXiv cs.CR

• StepJack benchmarks multi-step indirect prompt injection in computer-use agents Researchers introduced StepJack to test multi-step indirect prompt injection against computer-use agents.

  • Computer-use agents that browse and interact with web environments are affected.
  • The attack distributes an adversarial goal across multiple innocuous-looking instructions.
  • The instructions are planted across a chain of pages in the agent's navigation path.
  • StepJack contains 480 benchmark test examples for evaluating agent safety. 📄 Source: arxiv.org · 📎 Coverage: github.com · 👁 via arXiv cs.CR

• Researchers develop LLM-driven fuzzer for JavaScript engines in PDF readers Researchers introduced PDFuzzer, an LLM-assisted fuzzer for PDF reader JavaScript engines.

  • PDF readers with embedded JavaScript engines are in scope.
  • Existing fuzzers often test individual PDF JavaScript API calls and may miss sequence-dependent vulnerabilities.
  • PDFuzzer uses an LLM to build grammars and infer API relationships from manuals and execution traces.
  • A constraint solver generates concrete API call sequences for fuzzing. 📎 Coverage: arxiv.org · 👁 via arXiv cs.CR

• LoRAScan detects backdoor prompts in LoRA adapters via activation spikes LoRAScan detects backdoor prompts in low-rank adapters for large language models.

  • The research applies to large language models using compact, potentially untrusted LoRA adapters.
  • Backdoored adapters can trigger harmful content, malicious code, political propaganda, or covert advertisements.
  • Hidden input triggers activate the malicious behavior while normal inputs appear benign.
  • LoRAScan identifies backdoors through spikes in down-projection activations.
  • Merging adapters with base models can dilute backdoor signals and reduce detection performance. 📎 Coverage: arxiv.org · 👁 via arXiv cs.CR

• SynChain Shows Computer-Use Agents Can Build Persistent Attack Chains Researchers demonstrated that computer-use agents can create poisoned artifacts that reactivate in later workflows.

  • The finding affects computer-use agents, including OpenClaw, Codex, and Claude Code.
  • SynChain embeds dormant malicious influence in persistent skills and memory entries.
  • Agents are induced to generate benign-looking artifacts that survive internal state updates and vetting.
  • The payload reactivates as trusted context in future workflows without new external input.
  • Researchers evaluated the technique using CUAChain, a dataset of 30 benign task chains and three attack objectives. 📄 Source: arxiv.org · 📎 Coverage: arxiv.org · 👁 via arXiv cs.CR

• gnss-spoofing-uav-swarms — arXiv cs.CR

• rag-poisoning-detection — arXiv cs.CR

🔓 CVEs & KEV

• CVE-2026-19387 — CVSS 7.6 — Gstreamer: gstreamer1-plugins-bad-free: gstreamer: heap out-of-bounds write i...
• CVE-2026-19389 — CVSS 7.1 — Gstreamer: gstreamer1-plugins-ugly-free: gstreamer: integer overflow/underflo...
• CVE-2026-72522 — CVSS 6.2 — libexpat before 2.8.3 has an out-of-bounds read and resultant infinite loop b...