One AI layer runs the programme.
You approve the change.

Every tenant is hard-partitioned at the storage and indexing layer. There are no cross-tenant queries and no shared indexes. Every AI job is scoped to a single tenant with the tenant id asserted at the boundary, and output that references another client is rejected before it is stored or shown. The MSSP cockpit only ever surfaces metadata across tenants, never raw findings, log lines or evidence. We can walk you through the tenancy boundary before you commit.

The platform runs in eu-west-2 (London). Raw and normalised event storage, backups and audit trails stay in region, except for the processors named in the privacy policy. Content is redacted before it reaches a model. Models are frontier models from Anthropic (Claude), called through Penthropic's own guardrailed client.

It triages every new alert as it arrives, runs a daily analyst sweep, opens and works a case for each incident, reviews evidence, checks your vendors' answers against the documents they uploaded, and drafts a monthly report. It also proposes remediation. It never applies one. Every proposal, decision, tool call and outcome is written to a hash-chained, tamper-evident log, and the per-job token use and cost are recorded alongside the decision.

Detection and AI investigation take roughly 1–2 weeks for connectors and rule-pack tuning. Sentinel is a single appliance image deployed inside your estate, outbound-only, with no inbound port. PhishSim and awareness training take under a week to the first campaign. The Penthropic endpoint agent is optional: it runs on macOS and Windows, is outbound-only, and collects only approved posture data such as installed software, browser extensions, listening ports and patch state. Most signal comes from log sources and cloud APIs you already have.

Three sources connect and are polled every five minutes: CrowdStrike Falcon detections, AWS CloudTrail through a cross-account role you create, and the Okta System Log. Connect CrowdStrike and its host inventory and Spotlight vulnerability data also land in your asset register every hour, and a connected AWS account is assessed for S3, IAM, security groups, CloudTrail, GuardDuty and Security Hub.

Anything else you run pushes events to a token-authenticated endpoint you create yourself, as JSON over HTTPS. That is the answer for a tool we have never heard of, and it means you are not waiting on us to write a connector. We do not accept syslog today, and there is no file drop.

Going the other way, findings can be raised as Jira issues and alerts delivered to a Slack channel. Where a source is one we cannot read yet, we say so by name rather than listing it as supported.

Both. Most clients start managed: we operate the platform, your team has access, and we handle tuning. You can move to self-managed at any point in the same tenant. There is also a split where we operate detection and you operate response. We recommend the shape that fits your team rather than the one that bills most.

Your data is yours. On termination we provide a full export: events, findings, evidence and the audit trail, in JSON, CSV or PDF. After your export is acknowledged, the tenant is destroyed within 30 days and a destruction certificate is issued. No lock-in clauses, no data-egress fees.

No. Advisory engagements stand on their own: fixed-fee assessments, third-party risk design, threat modelling sprints and fractional vCISO. See the advisory page. The platform is there if you want continuous operations after the project, not a condition of working with us.