Latest in AI
Whisper Graph XDR: AI Agent Security, Network Controls and Pricing
Whisper gives agents a network identity and an external cutoff mechanism, with important differences between DNS monitoring and full traffic enforcement.
Whisper Security launched Graph XDR on October 28, combining endpoint detection, network identity and external containment for machines and AI agents. Its central claim is that security teams can cut off a compromised identity from Whisper's network rather than rely entirely on software running inside the compromised host.
That is a useful distinction for agent infrastructure. An application can tell an agent which tools it should use, but operators also need a way to restrict where its traffic can actually go.
How Whisper Graph XDR works
The launch announcement describes an individual public IPv6 address for each participating machine, container, device or agent. Whisper checks destinations against its infrastructure graph and lets operators revoke an identity from outside the host.
The product overview describes three connection options: a tunnel for routed traffic, an egress proxy for selected applications, or DNS-only integration. DNS translates names into network addresses; egress controls govern outgoing connections.
Those options do not provide identical coverage. A DNS policy can constrain name resolution, while routed enforcement can evaluate the connections passing through it. An organization evaluating the product needs to establish which paths are covered and whether a workload retains another route out.
Whisper also offers host sensors. Linux has a kernel-level enforcement path, Windows uses a user-mode sensor, and deeper macOS support remains in development. These are vendor-described capabilities, not independently established detection results.
What changes for AI agents
Whisper's engineering overview describes identities arranged in a tree, with an operator able to revoke a parent and the identities beneath it. The console shows the affected scope before confirmation.
For a build service that launches several task agents, that introduces a practical containment decision: stop one agent or stop the shared service and its descendants. The second choice may reduce exposure faster, but it can also interrupt unrelated work.
The network identity is different from an application login token. One establishes which service account can call an API; the other helps attribute and restrict traffic. Teams may need both. Neither tells them whether an agent's proposed code change or business decision is correct.
Whisper Graph XDR pricing
The published pricing page separates Graph XDR from Whisper's intelligence-query product:
| Graph XDR plan | Published price | Important detail |
|---|---|---|
| Free | €0 | Five identities, one seat and seven-day retention |
| Team | From €3.99 per endpoint per month, billed annually | €4.79 with monthly billing |
| Business | From €5.99 per endpoint per month, billed annually | €7.19 with monthly billing |
| Enterprise | Custom | Contract-based deployment and support |
One IPv6 identity is billable whether it belongs to a laptop, server, container or AI agent. Paid plans require a conversation with Whisper; the page says there is no self-serve checkout yet.
The detailed comparison also matters: Free includes a manual off-host cutoff but does not include the paid egress-control capability. A free test therefore should not be treated as proof that the complete paid enforcement setup fits a production workload.
The limits to test before deployment
Whisper says resolution fails open by default if its graph is unreachable, with a strict blocking mode available. That presents an operational choice: preserve connectivity during a dependency failure, or preserve the stricter network boundary and accept possible interruption.
The pricing page also lists SSO/SAML and external alert routing on its roadmap. Buyers should check those requirements before assuming the product fits an existing security operations process.
No network cutoff can undo data already sent, and a trusted destination can still receive an inappropriate request. Graph XDR should be evaluated as a containment layer, with tests for bypass paths, revocation timing and disruption to legitimate traffic. The important outcome is a boundary that keeps working when the agent stops behaving as intended.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders explain AI infrastructure and security products clearly, so their launch stories communicate both the mechanism and its limits.