Supply-chain security

Use agentic AI without importing someone else’s risk.

Skills and MCP connectors make AI useful at work. Harriet gives IT the risk signals, containment, and approval evidence to say yes without giving untrusted packages a free pass to company data.

Connector review Before activation
Payroll helper MCP Third-party package · sensitive access
Risk signals Review needed
Decision trail Approved with evidence
Reviewer: IT admin Audit-ready
The problem

Agentic AI has a supply chain now.

A helpful skill is still instructions your assistant may follow. A useful connector may still be third-party code with access to company systems. The question for IT is not whether teams should extend AI. It is how to approve those extensions without creating a new blind spot.

Unreviewed instructions become policy

A shared skill can quietly tell an assistant what to trust, what to reveal, and which safeguards to ignore.

Connectors expand the blast radius

MCPs can reach documents, credentials, internal systems, and personal tokens. That access needs a review trail before it is granted.

Approved once is not trusted forever

Packages, branches, and third-party sources can change after approval. Security needs evidence that the thing running today is still the thing you reviewed.

Shadow AI becomes shadow software

Without a gate, every enthusiastic team can add a new mini supply chain faster than IT can see it.

What Harriet changes

Security gets a decision layer, not another queue of exceptions.

Harriet keeps the public promise simple: risky additions are reviewed, contained, blocked when necessary, and auditable later.

  1. Stop risky additions before data access

    Harriet turns each submitted skill or connector into a reviewable decision, with risk signals surfaced before it can touch company systems.

  2. Give reviewers a clear yes, no, or fix path

    Security findings land beside the approval workflow, so reviewers can approve clean items quickly and send risky ones back with context.

  3. Contain third-party execution

    Hosted connectors run away from Harriet application servers, with access scoped and activity recorded for later review.

  4. Walk into audit with evidence

    Every approval, override, scan, and tool call leaves a trail, so compliance review is not a forensic exercise months later.

Why security teams care

Move from “who installed this?” to “here’s why we approved it.”

Harriet gives every extension a paper trail before it becomes part of the way your company works. That means faster approvals for clean requests, fewer awkward exceptions, and less scrambling when compliance asks what changed.

  • SOC 2 Type II context
  • EU data residency available
  • Human approval before activation
  • Append-only activity evidence
Signals, not secrets

Enough proof to build trust, not a recipe to copy.

The review surfaces the categories that matter to a buyer — provenance, access, unsafe behaviour, containment, and auditability — without publishing the exact scoring recipe.

Risk scoring before activationHuman review queueOwner override evidenceSandboxed connector executionOngoing drift checksTool-call audit history
The line we draw

We focus on the extensions your team controls.

Model vendors still matter, and we cover AI requests through contracted sub-processors, DPAs, and no-train commitments. This page is about a different risk: the skills, packages, and connectors your own organisation chooses to add on top.

Bring the connector you would hesitate to approve.

We’ll show how Harriet turns it into a reviewable, contained, auditable decision.