ARKLINKIndependent AI Security & CyberOps research lab

Trust isthe attacksurface.

Security tooling for AI agents, MCP servers and cyber deception. We watch what attackers actually do - not what scanners report.

Work with us View research

Track record - every number links to its record

001The new trust boundary

Autonomous agents read tool descriptions, decide what to call, and act - faster than any human reviewer, with broader credentials, and almost no audit trail designed for what they just did. Traditional security tooling was built for humans behind human-paced UIs. That model is collapsing. We build for the boundary that just moved.

P/01

Exploitability over noise.

A finding that is not reachable is not a finding. We optimise for issues a real attacker can actually use.

P/02

Deception as ground truth.

If a real attacker touched it, it matters. Telemetry from honeypots beats signature heuristics.

P/03

Research, not theatre.

We publish methodology, not vibes. Findings come with reproductions, corpora, and provenance.

002Flagship - open source

honeymcp

An open-source honeypot for MCP-native AI agents. One Rust binary - 15 MB, SQLite on disk, 256 MiB of RAM.

Speaks Streamable HTTP (MCP spec 2025-06-18) and legacy HTTP+SSE side by side, ships four production personas, and tags every request with MITRE ATT&CK / ATLAS technique IDs at write time. Releases are SLSA Level 3 build-provenance signed and exportable to STIX 2.1.

7Threat detectors mapped to MITRE ATT&CK / ATLAS

4Production personas - Postgres · GitHub · Vercel · Stripe

View on GitHub

honeymcp · attack storycaptured

persona: github-admin

POST /mcp · streamable-http · spec 2025-06-18

5 events

initializeclient unknown-agent/0.4 · MCP-Protocol-Version: 2025-06-18
+0.0s
tools/list - recon + tool_enumerationread 14 tool defs · detectors: T1592, AML.T0040
+1.4s
prompt_injection - tool description echo“ignore previous” planted via tool description
+12s
secret_exfil - tools/call get_envtarget AWS_SECRET_ACCESS_KEY · CVE-2025-59536-class hook injection co-fires
+24s
session closeSTIX 2.1 export ready · ATT&CK / ATLAS tagged at write time
+97s

SLSA-3 · cosign-keyless · SBOMSTIX 2.1 · ATT&CK · ATLAS

003Research index - verified work only

01R-01 · FLAGSHIPMCP deception & threat intelhoneymcp · 7 detectors · STIX 2.1 02R-02 · OPEN SOURCEKernel fuzzing infrastructureProject Zero SockFuzzer · aarch64-darwin 03R-03 · CVE · LINUX KERNELIPsec state use-after-freeCVE-2026-46116 · own patch · mainline + 7 stable trees 04R-04 · CVE · APPLE XNUApple kernel memory safetyCVE-2026-43822 · CVE-2026-43739 · iOS / macOS 26.6 05R-05 · ACTIVEAI agent security reviewPrompt injection · tool misuse · RAG abuse 06R-06 · EXPLORATORYExploitability-first triageReachability · scanner-noise reduction 07R-07 · BOUNTY · 2× PAIDPublic bug-bounty programsOpenAI (2 awarded) · HackerOne - 8 reports / 4 programs

004Work with us

The same people who did the research above, on your system. Scope and price agreed in writing before work starts.

S/01AI agent and MCP security reviewTeams shipping agents, copilots or MCP servers into production.S/02MCP deception with honeymcpSecurity teams who want to see what attackers do to agent infrastructure, not guess.S/03Exploitability-first triageTeams drowning in scanner output who need to know what is real.S/04Kernel and systems fuzzingVendors and platform teams who own kernel, driver or network stack code.

Start an engagement All services

005Start a conversation

We work with teams building AI agents, MCP servers and security products that need high-signal research - not generic checklists.

hello@arklink.co

Start an engagement Report a vulnerability

Founder: Michał Kosiorek
Location: Poland · EU · Remote
Registry: CEIDG - NIP 9522098675
Existing clients: Open a support request →