OpenA2A "AI Agent Security Posture" report: tool calls showed up

The 6th issue of the OpenA2A Research "AI Agent Security Posture" behavioral threat report shows that its TrapMyAgent honeypot MCP server received 420 tool calls from September 19 to October 19, 2026, a large increase from 9 in the previous window, while connection events dropped from 422,010 to 396,401. The calls came from 19 fingerprints and involved operations such as shell execution, database queries, and secret acquisition. Of the 2,813 injection decoy callbacks recorded by the AgentPwn honeypot, 2,578 came from browsers and link prefetchers, 235 came from LLM crawlers that declared their identity, and 0 declared themselves as autonomous agents. A2A handshakes rose from 15,670 to 27,247, with only 1 task request after the handshake, which the report judged as enumeration rather than use; the October 19 scan also found 2,495 exposed MCP servers, an increase of 24.9% over September 1.