Agent security is product security, and your marketing site counts
A marketing site used to have two kinds of visitors: people and search crawlers. Now an agent can build a The post Agent security is product security, and your marketing site counts appeared first on The New Stack .
A marketing site is not just a collection of pages on the web - it can be seen as a product that is built, read and optimized by various agents, who can all have an impact on the site's integrity. These agents include builders, readers/scrapers, analyzers, optimizers and adversarial agents that attack the site. While the site may be public with no login required, it still needs to protect three kinds of integrity: content, delivery and signal.
Builder agents are the first agents adopted by a team and set the pattern for others. They can write pages, config, scripts and redirects based on inputs like docs, tickets, competitor pages or CMS fields. The risk lies in what these agents read before they write. Examples of recent incidents show the potential consequences, such as AWS removing a destructive prompt from its Amazon Q Developer extension after it was shipped to production, or Replit's agent accidentally deleting a production database during a code freeze.
To protect the site, three layers around the agent are recommended: the agentic SDLC process, the agent harness control layer, and the sandbox containment layer. The SDLC process outlines where agents act across plan, draft, review and publish, with different stages requiring human decisions. The harness control layer decides which tools the agent can call, with which credentials, and what gets logged or approved.
The sandbox layer provides an isolated runtime for the agent with no standing credentials, restricted network egress, and no path to production data.
For a marketing site, this translates into giving builder agents draft access and keeping publishing behind a person, reviewing a diff of content, scripts and redirects, treating every input as untrusted and adding provenance records. Webflow's AI site builder, for example, connects AI tools like Claude to a site through an MCP server, allowing an agent to design, build and manage the site through natural conversation or agentic workflows.
Webflow's controls map to these recommendations, with agent instructions guiding the agent based on design system, voice, CMS and policies, while ensuring only authorized users can change them.
Written by urgent.news from The New Stack's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.