MCP rug pulls and Agent Card poisoning threaten A2A protocols

Researchers detailed supply-chain attacks on AI agent fleets: MCP servers mutate tool metadata after approval, while unsigned A2A Agent Cards inject hidden instructions. In 1,000 simulated attacks, compromise rates reached 94.2%, while cryptographically pinned schemas blocked 100%.

• tools/list rug pull compromised 94.2% of unprotected agents
• Poisoned agent.json in A2A yielded 91.8% compromise rate
• Cryptographic schema pinning blocked 100% of attacks in 2.1 ms
• CVE-2025-6514 and CVE-2025-6515 confirmed dynamic metadata risk