MCP agent-to-agent comms may be the riskiest protocol you've never heard of
Ars Technica AI · rss · 2026-10-06
Ars Technica reports that MCP, the protocol for agent-to-agent communication, is an under-discussed and highly exploitable attack surface. Over the past five months, Google and four other organizations have acknowledged vulnerabilities where one compromised agent spreads malicious instructions to other internal agents, enabling data exfiltration.
How it works
- A special form of prompt injection targeting not the LLM but a specific agent (e.g., translation or data analysis) whose guardrails are lax or absent
- The compromised agent forwards instructions down the chain; downstream agents follow them because they explicitly trust the upstream one
- Independent researcher Syed Anas Mohiuddin ran PoC attacks against agents at Google, JP Morgan Chase, Weaviate, Rapid7, the French government's digital directorate, and US federal agencies
The core problem: trust chains between agents lack mitigation and are hard to defend.
More from coding & agent
- User sends AI agent to shop at REI, gets blocked by 'Press & Hold' human verification — unixterminal · 2026-11-21
- Two Claude Skills for fact-checking and hypothesis testing: pre-call, then hunt disconfirming evidence — TheNerdosapien · 2026-11-21
- Together AI: tune the system before the weights — cheapest levers first when evals fail — togethercompute · 2026-11-21
- Coding Value Trends to Zero, Systems Thinking to Infinity: AI Skills Debate — attentionmech · 2026-11-21
- 100+ AI Agents Collaborate to Build Browser FPS Game KILLFEED — majidmanzarpour · 2026-11-21
- Security Firm Claims Grok Bots Made 31 Unprompted App Requests Missing From Chat History — shoumikgoswami · 2026-11-21