A procurement agent and a recruiting agent browse the same web and need opposite permissions. One reads pricing and case studies; the other reads careers pages. Neither should ever reach a login form. These fifteen pages work through the real jobs, one at a time.
Each page below takes one agent job and answers three questions: which page types that job legitimately needs, which ones must stay denied even on domains the agent is allowed to visit, and what the policy looks like written down.
The pattern repeats across all fifteen. An agent's job describes what it should read. Risk lives in what it can do — sign up, upload, post, spend. Page types separate the two on the same domain, so a policy can say yes to stripe.com/pricing and no to stripe.com/login without the blunt instrument of blocking the domain.
Agents that gather evidence about vendors and markets. They read deeply and widely, which makes their allow list broad — and makes the handful of denied page types matter more, not less.
How to let procurement agents compare vendor pricing and case studies without ever reaching a checkout page. Policy walk-through and a YAML template.
Mining about, leadership, legal and security pages plus jurisdiction data, while account creation on the prospective vendor's own site stays denied by default.
Verified blog and press URLs across 40M+ domains let a research agent sample by popularity rank instead of guessing paths or walking into login pages.
Verified documentation and case-study URLs instead of guessed paths, with the one-lookup token-cost math worked out in full.
The one workflow where checkout access is allowed — with human authorization. How to design the approval gate, spend limits and the audit trail around it.
Agents pointed at other companies on your behalf. Here the risk is not only credential exposure — it is an agent acting in your name on someone else's site, and the brand damage that follows.
Mining leadership, careers and case-study pages for account data, and why the same domains' login pages must stay denied by default.
Read a prospect's blog, docs and pricing freely — but never let the agent submit its contact form, post a comment or sign up for its newsletter.
Building an enrichment pipeline on verified contact, about and leadership URLs, plus domain metadata for dedupe — not on guessed paths.
Verified careers-page URLs across 40M+ domains rather than guessed /jobs paths, plus the contact-page etiquette rules a sourcing agent must never cross.
Agents that revisit the same URLs on a schedule. Verified page-type lookups replace a re-crawl on every run, which is where most of the cost and most of the failures sit.
Why price monitors should stop re-crawling for the pricing page on every run, how verified URLs cut the failure rate, and how to set a per-domain refresh cadence.
Watching docs, changelogs, status and pricing across a competitor set, and the ethics boundary that keeps the agent out of login-gated areas.
Why guessing status.domain.com fails, and how verified status-page URLs fix uptime and third-party dependency monitoring.
Reaching blog and press pages directly, and routing around paywalled login surfaces cleanly instead of hammering them.
Agents that operate inside a process with an auditor at the end of it. The log of what was denied matters as much as what was allowed.
An agent that reads vendor help centers, docs and status pages across 40M+ domains, and never touches a login, account or ticket page.
Verified legal and privacy URLs for tracking terms changes, and a full audit trail of every navigation decision, allowed or denied.
Several high-profile cases last year involved AI agents escaping their task or being steered off it — the Hugging Face breach, the DseWiki hijack, a covert channel through JFrog Artifactory, and a run of third-party account takeovers. Our analysis shows the documented attack steps in almost all of them would have been denied at the request, by the page-type database and the egress rules together.
Would your agents have been stopped? Check the incident analysisNot sure which page types your job actually needs? The page-types database lists all 28 with definitions, and the free sample CSV shows 100 real domains in the production schema. For humans rather than agents — controlling which AI tools staff can reach — that is a different product: aitoolsblocklist.com.
Once you know which page types your agent needs, the next question is where the check runs and who signs off on it.
LangChain, the OpenAI Agents SDK, MCP servers, Playwright, computer use — where the policy hook goes in each.
Industries — 12Banking, insurance, government, ecommerce, and the vendors who ship guardrails to their own customers.
Learn — 12Why domain blocking fails, what a credential surface is, and why a prompt instruction is not a control.
