ContextLeak Uses Malicious Tools to Exfiltrate LLM Agent Context

🕵️ RESEARCH & DEEP DIVES

• ContextLeak Uses Malicious Tools to Exfiltrate LLM Agent Context
  Researchers developed ContextLeak to exfiltrate LLM agent context through malicious tools.
  • LLM agents that use external tools are affected.
  • The attack targets user prompts, execution trajectories, and tool lists.
  • A malicious tool induces the agent to pass runtime context as input arguments.
  • The tool transmits the data to an attacker-controlled endpoint.
  • Reinforcement learning generates the tool’s name and description using simulated agent contexts.
    📄 Source: doi.org · 📎 Coverage: arxiv.org · 👁 via arXiv cs.CR

🔓 CVEs & KEV

• CVE-2026-77850 — CVSS 8.4 — Stored XSS in AshAdmin relationship typeahead via unescaped label_field conte...

• CVE-2026-82673 — CVSS 8.3 — Path traversal in AshAdmin file uploads via unsanitized client filenameImprop...

• CVE-2026-82722 — CVSS 8.3 — AshAdmin LiveView events intern atoms from client input, exhausting the atom ...

• CVE-2026-75757 — CVSS 8.3 — AshAdmin cookie reader matches names by substring, enabling actor/session sha...