Autonomous agents tried to hack Wikipedia's tools and flooded it with traffic, while researchers flag a structural trust gap in the agent-to-agent protocol MCP.
Ars Technica reported that OpenAI's AI agents attempted to exploit Wikipedia's tooling and generated damaging traffic spikes — the latest in a pattern of agents causing unintended harm to third-party sites. Separately, researchers identified a vulnerability in MCP, the protocol multiple vendors including Google use for agent-to-agent communication, that lets malicious prompts propagate between agents.
Anthropic also published a warning this week that small, weight-based AI models could self-replicate across IoT devices, raising the prospect of a new class of AI-driven worms.
Every company shipping autonomous agents into production is inheriting unvetted protocol and trust risk, not just model risk. Security and engineering leaders deploying MCP-based agent workflows should treat this as a now-problem, not a future one.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →