Foxhound

AI-enabled security assessment, with a human at the trigger.

Foxhound automates approved assessment steps within your scope. A tester approves active testing, investigates the results, and signs off findings before delivery.

Scope an engagement Open the portal
The Foxhound findings portal An abstracted view of the Foxhound portal: a list of properties under assessment on the left, and a findings table on the right showing severity, CVSS score and remediation status for each finding. Finding titles are redacted. foxhound.fenko.nz/portal Properties 12 8 3 0 Burrow connected Findings live all open retest Severity Finding CVSS Status CRITICAL 9.1 OPEN HIGH 8.2 OPEN HIGH 7.4 VERIFIED MEDIUM 5.6 RETEST LOW 3.1 FIXED Severity distribution 4 critical 9 high 13 medium 8 low 3 info

Pentesting hasn’t kept up with how you ship.

You deploy every week. The assessment happens once a year, and what you get back is a static PDF. No live view while the work is happening, no retest loop when you fix something, no trail you can query afterwards.

The model

The agent does the work. The human owns the judgement.

Automation handles the approved routine steps. Testers own the approvals, the verification, and the manual depth on top.

Approve

Before anything touches the target

Scope, rules of engagement and every active check are signed off by a tester before they run. Nothing probes a host that was not on the list.

Verify

Before any finding reaches you

A tester reproduces a finding and writes its remediation guidance before it reaches you. Anything that does not reproduce does not get delivered.

Methodology

Six phases. Each one a gate, not a guess.

Automated steps follow published methodology, so coverage stays consistent between engagements rather than depending on what a tester happened to recall.

Assessment phases and human gates 6 phases run in sequence: Setup, Recon, Scan, Exploit, QA, and Report. A human gate labelled Approve sits in front of Exploit. A human gate labelled Verify sits in front of Report. Setup Recon Scan Approve Exploit human QA Verify Report human

Measured against

  • OWASP WSTG v4.2
  • OWASP Top 10
  • OWASP API Top 10
  • OWASP LLM Top 10
  • PTES
  • NIST SP 800-115
  • NZISM v3.9
  • ISO 27001:2022
  • SOC 2 Type II
  • PCI DSS v4.0

What you get

A live portal, and a report a senior consultant would sign.

The portal

Findings land as they are confirmed, not at the end. Critical and high severity issues are flagged the day we verify them, so your team can start remediation while the assessment is still running.

  • Severity scored with CVSS 4.0 and mapped to CWE
  • Evidence attached to every finding: requests, responses, reproduction steps
  • Retest on request, with the status updated in place
  • Viewer, manager and admin roles across all your properties

The report

A branded PDF with an executive summary in plain language for leadership, and a technical section your engineers can work from. Download it as often as you need.

API and MCP

A scoped REST API and a Model Context Protocol endpoint. Pull findings into your own systems, or query engagement status from your AI tooling.

Retesting included

Fix something, request a retest, and we confirm it and update the finding. There is no separate invoice for that.

Tailoring

Foxhound, scoped to your environment.

There is no separate bespoke service to buy. Foxhound carries the automated work, and the parts of an engagement that need a human decision are where we shape it around your environment.

Scoped to your environment

Before the first run

We work through your architecture, your threat model, and what is genuinely off limits, then encode that as the scope the agent runs against. Blackbox, greybox or whitebox, whichever actually tells you something.

Manual depth on top

Where the agent stops

Business logic flaws, chained vulnerabilities and access control gaps that only surface once someone understands how your application is meant to work. A tester goes after those directly, in the same engagement, and the findings land in the same portal.

Your own instance, if you need it

Regulated environments

Run on Fenko’s managed platform, logically isolated per engagement, or on a dedicated deployment with no shared compute or data plane where data residency requires it.

Vulnerability research

What it turns up in real software.

Foxhound is pointed at client scope, but the same passes surface bugs in the software underneath. Some of this we found. Some of it we verified for the vendor after another researcher reported it. Each card says which.

  • CVE-2026-65819 CVSS 7.5

    Eleven layer decoders panic on crafted packets

    Target
    gopacket < 1.7.0
    Class
    Out-of-bounds read

    Decoders across TLS, DHCPv4, sFlow, IPSec AH, VRRPv2, Diameter, GTPv1-U, ERSPAN II, LCM, RadioTap and Dot11 read or slice packet data using attacker-controlled length, count and offset fields before checking them against the buffer. A malformed packet panics the decoding goroutine, reachable through DecodingLayerParser or a direct DecodeFromBytes.

    Credited to Fenko. Published 27 July 2026 as GHSA-8mcr-459q-5mx2, fixed in 1.7.1. Read the advisory
  • CVE-2026-54345 7 Jun 2026

    Diameter AVP decoder: uint32 underflow on vendor header size

    Target
    gopacket <= 1.6.0
    Class
    CWE-191, CWE-770

    An integer underflow on the vendor header size lets a single crafted message drive an unbounded allocation of roughly 4 GiB, exhausting memory without authentication.

    Reported by tonghuaroot. Fenko is credited as remediation verifier: we confirmed the fix, we did not find the bug. Published 7 June 2026, fixed in 1.6.1. Read the advisory
  • CVE-2026-54332 7 Jun 2026

    sFlow decoder: unbounded allocation from a 104-byte datagram

    Target
    gopacket <= 1.6.0
    Class
    CWE-770

    A 104-byte UDP datagram reaches an attacker-controlled make in the ExtendedGatewayFlow decoder and can request up to 16 GiB, exhausting memory without authentication.

    Reported by tonghuaroot. Fenko is credited as remediation verifier: we confirmed the fix, we did not find the bug. Published 7 June 2026, fixed in 1.6.1. Read the advisory
  • CVE-2026-12606 7 Jun 2026

    HTTP request smuggling via header-name parsing

    Target
    Eclipse Grizzly 2.1.11 to 5.0.1
    Class
    CWE-444

    Foxhound flagged a grizzly/2.4.4 banner on a pre-auth endpoint. The parser ends a header name at the first colon without checking for CR, LF or space, so a Content-Length tucked into a header name never gets treated as framing.

    Same defect and same fix as CVE-2026-12606, which Eclipse assigned from a separate trailer-section report. Our analysis reached security@eclipse.org on 7 June 2026 and reproduces back to 2.1.11, wider than the 4.0.0 to 5.0.1 range published against that ID. Read the research

Further advisories are in coordinated disclosure and will be listed here once they are published.

Where the work lands

  • Jira
  • Linear
  • GitHub
  • Slack
  • Teams
  • ServiceNow
  • SIEM
  • Webhooks

Safe by design

Autonomous, not reckless.

The agent works from an approved scope and has no way to widen it on its own. Out of scope means stop, not warn.

  • Proof of concept is read-only. Nothing is modified or exfiltrated.
  • Active testing runs against approved hosts, inside an approved window.
  • A tester can halt an engagement at any point.

In context

Speed and breadth, without losing the rigour.

Measure Annual consultant DIY scanners Foxhound
Speed Weeks Minutes, noisy Hours, verified
Output Static PDF Raw tool output Report and live portal
Coverage What one tester knows Whatever the signatures cover The full methodology, every run
False positives Filtered by the tester Yours to triage Filtered before delivery
Retest Scoped and quoted again Run it yourself Requested in the portal, included

What Foxhound assesses

Web applications, REST and GraphQL APIs, mobile apps on iOS and Android, internal and external network infrastructure, cloud configuration across AWS, Azure and GCP, source code, and AI systems including prompt injection and agent tool-use testing.

Full coverage is described on the penetration testing page.

Scope an engagement.

Tell us what you want assessed and we will quote against the scope. Pricing is per trigger, not per URL or per IP, because those penalise a fair account of your own attack surface.

Get in touch Open the portal