What we sell against these gaps
9 paid products on redhub.ai, one for each of the nine domains this assessment measures. That is an unusually complete slate and it is worth being suspicious of rather than pleased about: an assessment whose domains happen to be the shape of a product catalog would be a sales form. These domains came from the clock the assessment runs on, and the mapping fell out afterwards. Where a future domain has nothing behind it, this page will say so, the way the governance property's does.
None of them reaches into a system of yours or acts on your behalf. Seven grade a description you write down; two, the reliability harness and the admissibility engine, read traces and proposed actions your own agent produces. All nine are deterministic and end in a verdict you can recompute rather than a judgment call.
RedHub AI publishes this site and sells these products. We say so here rather than presenting them as neutral recommendations.
None of these is a penetration test and none of them certifies anything. They are structured procedures over a description you supply, and buying one changes nothing about what your agent can reach.
The nine
Which one closes which gap
Scores every connector, MCP server and OAuth grant on scope, data sensitivity, write capability and staleness, then takes the worst connector as the fleet's headline rather than the average. An unowned connector that can write regulated data -- personal, health, payment or otherwise legally restricted -- is ungoverned whatever its score.
Six controls per identity class across the service accounts, keys, tokens and agent credentials -- the accounts nothing creates a joiner or leaver event for, so their number is not knowable without enumerating them per system. The verdict is the weakest control, never the flattering mean, and unvaulted secrets together with no working revocation path is dispositive.
A pre-install trust decision on a third-party MCP server or skill, on six signals, not a code scan. Credential handling and update behavior together are dispositive, which is the clean-then-poisoned update pattern.
Checks the actions an agent proposes against your own domain rules before any of them run. It returns no admissibility score, because a contradiction cannot be scored.
Grades what each of your stop controls has been shown to do, not what it claims, returning STOPS, SLOWS or NAMED ONLY per control. It is the product behind one of the two routes back the gate on your result reads, and it draws the same distinction that gate turns on. Like this assessment, it reads each agent by its strongest control rather than its average.
Grades each action on whether you could undo it in time and on how many records one call touches, prints which of the two decided, and never treats a typical count as a bound.
The gap between an automated action and the record of it: whether the action and its record can come apart, and whether the record is usable once it is there.
Evaluates an agent at the trajectory level on six evaluators, task success first and then tool choice, argument validity, step efficiency, cost and policy, and gates a build on the verdict rather than reporting it.
The operator's go/no-go before an agent runs unattended, on five controls. Its teaching example is a refund agent at 86 out of 100 that is still DO NOT DEPLOY, which is where this assessment's own top band comes from.
Why the Agentic AI Security Bundle is not on this page. It exists, it costs $269, and it is cheaper than its four parts, which come to $316 separately. Two of the four are on the list above: the MCP Server & Skill Trust Gate and the Non-Human Identity gate, $158 together. The other two are the Indirect Prompt-Injection Exposure Gate and the Agent Memory & Context Poisoning Exposure Probe, which grade what a model reads and what it remembers.
That is a different question from the one this site asks, and this assessment never asks it. Recommending the bundle from a verdict here would be charging you $269 for an answer to something you did not come with. If what a model reads and remembers is also a live worry for you, the bundle is good value and the arithmetic above is the whole of it.
There is no subscription on this page either. The recurring product in this catalog maintains a governance posture, not an agent's guardrails, and offering it here would be selling the nearest subscription to hand.
How the assessment is scored is published in full, weights and bands included.