Microsoft Outlook RCE Vulnerability Allows Remote Code Execution

🕵️ RESEARCH & DEEP DIVES

• Study Maps How LLM Prompt Injection Reaches Classic Web Vulnerabilities Researchers have identified LLM-mediated web attacks that carry attacker input into traditional application vulnerabilities.

  • LLM-integrated chatbots, tool-calling pipelines, and agentic web applications are affected.
  • Attacker-controlled input can reach database queries, HTTP requests, file operations, template rendering, and API calls.
  • The study defines LLM2SQLi, LLM2XSS, LLM2SSTI, LLM2CommandInjection, LLM2IDOR, LLM2CSRF, LLM2XXE, and LLM2SSRF attack variants.
  • The LLM transforms attacker input before it reaches a traditional web-application sink. 📄 Source: arxiv.org · 📎 Coverage: undercodetesting.com · 👁 via arXiv cs.CR

• Study Finds Agentic LLM Security Research Trails Attack Research A review finds security research on agentic LLMs lags attack research.

  • The findings apply to autonomous LLM agents with real-world privileges.
  • Agentic systems combine planning, tool invocation, code execution, and persistent memory.
  • A compromised reasoning step can trigger unauthorized data access, irreversible state changes, or cascading failures.
  • Researchers screened 743 records and retained 85 papers published from 2023 through 2025. 📎 Coverage: arxiv.org · 👁 via arXiv cs.CR

• Researchers demonstrate programmable backdoor in vision-language models Researchers demonstrated a programmable backdoor in vision-language models.

  • The attack applies to vision-language models (VLMs).
  • A single poisoning phase implants a backdoor that enables arbitrary target-caption control.
  • Attackers select previously unseen caption semantics at inference time without retraining the model.
  • Stealthy triggers are synthesized on demand for the chosen captions. 📄 Source: arxiv.org · 📎 Coverage: arxiv.org · 👁 via arXiv cs.CR

• Devlore protects confidential VMs from malicious device interrupts Researchers introduced Devlore to protect confidential VMs from interrupt manipulation attacks.

  • Confidential VMs are designed to protect sensitive workloads from hypervisors, host operating systems, and co-resident VMs.
  • Malicious device interrupts can compromise confidential VM confidentiality and integrity.
  • Devlore delegates interrupt management to the hypervisor while trusted software performs correctness checks.
  • The Arm CCA prototype supported four devices and reported 0.06% overhead for typical integrated GPU applications. 📎 Coverage: arxiv.org · 👁 via arXiv cs.CR

📋 ADVISORIES

• Critical RHACM flaw enables GitOpsCluster controller token redirection 📄 Source: access.redhat.com

🔓 CVEs & KEV

• CVE-2026-70329 — CVSS 8.8 — Microsoft Outlook Vulnerability Allows Attackers to Execute Malicious Code Remotely
• CVE-2026-18961 — CVSS 8.1 — Social Login, Passkeys, Magic Link & Email OTP – Passwordless Login by Ventra...