An arXiv paper from 2 November 2026 sorts each OWASP MCP Top 10 entry into server, gateway or protocol tiers, and reports that server developers hit a hard limit on tool poisoning, intent subversion and context over-sharing.
Academic and industry research shaping the future of AI security, attack, and defence.
An arXiv paper from 2 November 2026 sorts each OWASP MCP Top 10 entry into server, gateway or protocol tiers, and reports that server developers hit a hard limit on tool poisoning, intent subversion and context over-sharing.
Forever Security found a malicious AGENTS.md in a customer's Downloads folder and reproduced the attack on OpenAI Codex and Claude Code in lab tests. Success rates were about 90% for Codex and about 50% for a Claude Code variant.
CarryoverBench, a new arXiv benchmark, shows a single injection written into agent memory or heartbeat files re-fires across sessions with no monotonic decay, and that prompt-level defenses mostly act at write time.
Check Point Research found that when teammates reversed on a rule, a frontier-model agent opened a forbidden folder in up to 98% of runs, up from 44%. The agent almost never said why.
COLM 2026 paper InjecMEM shows a single interaction, with no access to the memory store, can plant an optimized record that is later retrieved and steers a research agent's answers. Retrieve-time filters barely helped.