<style>
/* ---------- Primary button ---------- */
.button_bg {
background-color: #a83018;
transition: background-color 0.3s cubic-bezier(0.4, 0, 0.2, 1);
}
.button:hover .button_bg,
.button:focus-visible .button_bg,
.submit-wrapper:hover .button_bg,
.submit-wrapper:focus-within .button_bg {
background-color: #d24a2f;
}
/* ---------- Secondary button ---------- */
.button_bg.secondary {
background-color: var(--neutral--900);
border: 1px solid var(--neutral--500);
transition:
background-color 0.3s cubic-bezier(0.4, 0, 0.2, 1),
border-color 0.3s cubic-bezier(0.4, 0, 0.2, 1);
}
.button:hover .button_bg.secondary,
.button:focus-visible .button_bg.secondary,
.submit-wrapper:hover .button_bg.secondary,
.submit-wrapper:focus-within .button_bg.secondary {
background-color: var(--neutral--700);
border-color: var(--neutral--300);
}
</style>[fs-list-field]:has(> *),
.tag-component {
transition: background-color 0.25s ease, border-color 0.25s ease;
}
.fs-list-active .tag-component,
.tag-component.fs-list-active {
background-color: #d24a2f;
border-color: #d24a2f;
}Industry signal
>
80
%
of unauthorized agent actions through 2028 will be policy violations, not attacks.
Gartner
Industry signal
Every
3
months
the average F500 agent estate doubles.
Gartner
Industry signal
21
%
of enterprises have a mature agentic-AI governance model.
Deloitte, 2026
“
MCP and function-calling is the place where it is most valuable, and today it's uncovered.
Platform architect, top-ten global bank
“
I haven't seen yet this type of controls in the marketplace.
AI architecture lead, top-ten global bank
“
Agents created pathways in the runtime, not logged, not repeatable.
Head of AI & CTO, financial-technology firm
01
02
03
40
%+
of agentic AI projects will be cancelled by 2027, weak controls a named cause.
Gartner
82
%
of enterprises have unknown AI agents in their environment.
Cloud Security Alliance, 2026
150
k+
agents per F500 enterprise by 2028, up from fewer than 15 in 2025.
Gartner
Sources: Gartner · Cloud Security Alliance, 2026
In-path latency
<100ms
Layers
Heuristic, semantic, behavioural
Classification accuracy
97.5%
Coverage
Prompt, response, tool call, payload
Code changes
None
Behaviour that diverges from intent across a session, the failure mode you cannot write a rule for in advance.
Three layers, one record: heuristic, semantic that reads intent, and behavioural that tracks drift over time, judged in-path by purpose-built small language models.
Yes. The MCP and function-calling step nothing else controls is judged across prompt, response, tool call and payload.
97.5% classification accuracy across six small language models, with a verdict in under 100 milliseconds.
None, and no SDKs. It runs entirely in your stack.
See across every cloud, vendor, and team in days, not months.