Model Context Protocol (MCP)
An open protocol published in late 2024 that standardizes how AI assistants connect to external tools, files and services; it speeds up agents and widens their attack surface.
A language model on its own can only produce text. To act, read a database, send an email, consult a customer file, it must be connected to tools. Each connection used to be built case by case. The Model Context Protocol, published in November 2024 by the vendor Anthropic as an open standard, then widely adopted by other players in 2025, standardizes this connection: an MCP server exposes tools and data in a common format, and any compatible assistant can use them. The effect is like that of a standard plug: thousands of connectors appeared within months. This ease has a security downside. A malicious or compromised MCP server can describe its tools misleadingly to hijack the assistant's behavior; responses returned by a tool can contain hidden instructions, a form of indirect prompt injection; a connector with overly broad rights gives a manipulated agent access to sensitive data or actions. Third-party connectors also form a new, little-audited software supply chain. For insurance, the protocol shifts AI risk from content produced to actions executed: the question is no longer only what the assistant says, but what it can do, and with what rights.
Textbook case: a broker connects its AI assistant to its email and management software through MCP connectors. A booby-trapped email contains invisible instructions; the assistant, asked to summarize the inbox, executes them and sends the client list with renewal dates to an outside address.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.