An adversary,
under discipline.
KeroxLabs is a small lab building Kerox — a Rust-native, terminal-first, vendor-neutral autonomous red team. An orchestrator reads an engagement plan and works an objective the way an adversary would — recon, exploitation, privilege escalation, lateral movement, C2 — not the way a scanner does. Every live action is dry-run by default and gated behind a human. Built in the open, by hand.
In this
issue.
Five pieces — the orchestrator that runs an engagement, the Spearhead LLM agent, the discipline that gates every action, the forum we are opening, and the roster of agents behind it all.
An adversary, not a scanner.
Kerox is not a scanner that runs nmap and prints a report. An orchestrator reads an engagement plan, fixes on an objective, and works toward it through whatever path actually opens up — chaining reconnaissance, exploitation, privilege escalation, lateral movement, and C2 the way a real operator would. When a door closes it tries another. Findings are designed to feed a planned attack → defend → verify loop, so every result is something a defender can act on.
An LLM red team.
Spearhead is the agent pointed at the AI in the stack. It probes the things only a language model gets wrong — prompt injection, system-prompt leakage, guardrail bypass, tool-call exfiltration — and is designed to report every finding against the OWASP LLM Top 10 and MITRE ATLAS, so it lands in a framework defenders already use. It leads; the network agent follows it onto the rest of the attack surface, and the report agent turns the run into something a defender can use.
Discipline before the first packet.
Before a packet leaves the wire, Kerox writes the engagement down — Rules of Engagement, a ConOps, a Deconfliction Plan, and an OPPLAN mapped to MITRE ATT&CK — and then is built to refuse to step outside it. Every live action is dry-run by default and waits on an explicit human approval; nothing runs outside authorized scope. The whole thing is meant to read like a real operation, paperwork and safeties included — offense you could actually sign off on.
A forum, for people who run real engagements.
The Den is the slow, threaded forum we are building for people who do this for real — operators, red teamers, and the defenders on the other side of them. Pre-flight RoE arguments, ATLAS mapping threads, engagement postmortems, and the long debates about offensive AI that nobody else wants to host.
No articles. No engagement metrics. Just a room with the right people in it. Opening Q2 2026.
- RFCWriting a ConOps an agent can actually follow— POSTS
- ATLASMapping a tool-call exfil chain to MITRE ATLAS— POSTS
- TRADECRAFTKeeping evil-winrm sessions alive across a pivot— POSTS
- REVIEWWhat does a clean deconfliction plan look like?— POSTS
Specialist agents, run in a sealed lab.
The orchestrator does not do the work itself — it dispatches specialists. Spearhead leads on the AI; the network agent takes the conventional surface; the report agent turns the run into a deliverable. Each one drives real, interactive tools — msfconsole, sliver-client, evil-winrm — inside persistent terminal sessions, answering prompts the way a person would instead of scripting around them.
Everything is designed to run in an isolated Kali sandbox on its own operational network, walled off from the machine that drives it. Offense stays in the box.
- spearheadBUILDINGLLM / AI RED TEAM
Prompt injection, system-prompt leakage, guardrail bypass, tool-call exfil — mapped to OWASP LLM Top 10 and MITRE ATLAS.
- networkPLANNEDRECON · NETWORK
Maps the attack surface and works services and trust paths — recon, enumeration, and lateral movement on authorized scope.
- reportPLANNEDSYNTHESIS · REPORTING
Turns the engagement into a deliverable — narrative plus findings, mapped to MITRE ATT&CK and ATLAS, as Markdown, JSON, or SARIF.
- webPLANNEDWEB APPLICATIONS
The web surface — injection, access-control, and logic flaws — once the wedge and recon are solid.
An adversary with no rules teaches you nothing — the discipline is what turns an attack into an answer.
Read the code.
Break it.
Write back.
KeroxLabs builds in the open. Patches, exploits, and hard questions about doing offense responsibly — bring them. The bar is technical, the reply is fast, the door is unlocked.