Skip to content
hatchbase

Kittiwake

Raising Seed

A runtime guard for AI agents that checks every tool call against policy and keeps evidence for audits

4.3(4 reviews)AI safetyEdinburgh, United Kingdom2 followers

Overview

Problem

Companies are giving AI agents access to email, payments and internal databases, but the controls are mostly prompts asking the model to behave. A single poisoned web page or support ticket can instruct an agent to forward customer data or approve a refund, and afterwards security teams often cannot reconstruct which instruction led to which action.

Solution

Kittiwake sits between the agent and its tools as a proxy. Each tool call is checked against written policies, such as spend limits, data classes and allowed recipients, and against a classifier that spots injected instructions in retrieved content. Risky calls are blocked or sent for human approval, and every decision is logged with the inputs behind it.

Target market

Banks, insurers and software companies deploying agents with write access to real systems; agent security budgets are forming inside application security spend, estimated at US$8B a year.

Business model

Subscription priced by screened tool calls, starting at US$2,000 a month for 5M calls, with an enterprise tier for self-hosting and audit exports.

Demo and media

Open pitch deck (opens in a new tab)

Team

  • Isla MacLeodCEO and Co-founder
  • Kwame AsanteChief Scientist and Co-founder
  • Fergus LindsayHead of Engineering

Traction and funding

  • Paying customers7
  • Monthly revenueUS$34K
  • Tool calls screened per day2.3M
  • False-block rate0.4%
Current roundSeed, raising US$3.5MPrevious: Pre-seed US$1.1M (2024)

Reviews and ratings

Write a review
4.34 reviews
  • 5 stars2 reviews
  • 4 stars1 reviews
  • 3 stars1 reviews
  • 2 stars0 reviews
  • 1 star0 reviews
Sub-scores
Product
4.0
Team
4.5
Potential
4.3
  • Aisha Bello
    Customer

    Clear traces, but the approval queue got noisy early

    We put Kittiwake in front of the agent that books appointments and answers billing questions in our patient app. The trace for each blocked call is clear, and it caught a hidden-text injection in a support ticket during our pilot. But approval requests piled up for the first three weeks until we tuned policies, and the proxy added around 120 ms per tool call. We evaluated it in our own staging environment, since no live demo is offered yet.

    Reply from Isla MacLeod, CEO

    Our new policy simulator lets teams tune thresholds on past traffic before going live, which cuts early approval volume by about half.

    Helpful (0)Report
  • Hannah Lindqvist
    Investor

    The right team for a problem that is about to get urgent

    Isla built payment fraud controls before founding Kittiwake, and Kwame published on injection detection before this, which explains why the classifier handles payloads split across sources. Seven paying customers at seed, including two regional banks, is strong for a category this new. Pricing per screened call should scale naturally with agent adoption.

    Helpful (0)Report
  • Ravi Menon
    Industry observer

    Held up well when I tried to break it

    In a sandbox the team set up for outside researchers, I spent an afternoon trying indirect prompt injections through retrieved pages and documents: hidden text, instructions in image alt tags, payloads split across two sources. The classifier caught most of them, and the policy layer blocked the rest because the resulting tool calls broke spend or recipient rules. Layering policy on top of detection means a missed injection still runs into hard limits.

    Helpful (0)Report
  • Lucas Ferreira
    Customer

    Gave our security team a real answer on agent permissions

    We wanted our support agent to issue refunds, and security would not sign off without hard limits and an audit trail. Kittiwake's refund policy and approval queue got it approved in two weeks. The decision trace showing which input led to each block is what convinced our auditors.

    Helpful (0)Report

Testimonials

Coming soon. Testimonials are invited by the founder and shown separately from open reviews.

Updates

Coming soon. Follow Kittiwake to hear about product and funding updates when this launches.