The four hints
readOnlyHint(default false): the tool doesn't change anything.destructiveHint(default true): a change may destroy rather than add.idempotentHint(default false): repeating the same call is safe.openWorldHint(default true): the tool reaches things outside a closed set, such as the web.
A tool without annotations is treated pessimistically: not read-only, possibly destructive, not idempotent, open-world. ChatGPT, for example, asks you to confirm every tool not marked read-only.
Not a safety guarantee
The MCP project is clear that annotations may not describe a tool faithfully, that clients should treat them as untrusted unless the server is trusted, and that hard guarantees need network controls or sandboxing. See Is this MCP server safe?