Things I've built

GTFOBins for AI agents.

A reference of the techniques that turn an AI agent into an attacker's tool: from prompt injection, through access to MCP servers, to privilege escalation. Every entry says how to recognise the technique, how it works and what closes it.

• Sorted by attack surface and by what it gives the attacker
• Mapped onto OWASP LLM and Agentic AI, MITRE ATLAS and CWE
• Severity levels from Info to Critical
• A reference, not a ready-made attack tool

A File Inclusion scanner for the techniques older tools don't know.

A modular LFI/RFI scanner. Built because neither LFISuite nor kadimus sees PHP filter chains or two-phase log and session poisoning. The interface follows ffuf: the INCLUDE marker points at the injection spot.

• Eight modules, from file read to code execution
• PHP filter chains, log and session poisoning
• Built-in HTTP server for RFI attacks
• Auto-detected traversal depth and deduplicated results

The boring first stage of a pentest as a single pipeline.

Recon, enumeration and initial vulnerability detection run as one chain: open ports from nmap become targets for whatweb, sqlmap and INCLUDED. What comes out is an interactive report you can attach to an audit, not a terminal log.

• Five steps, each feeding its results into the next
• Findings mapped onto NIS2, UKSC, ISO 27001 and DORA
• Offline report with no CDN, severity filter and JSON export
• Docker image on Kali with the whole toolset inside

A CTF engine that holds up when hundreds of teams pile in.

A self-hosted platform for running Capture The Flag events: a Go server, a React interface, the whole stack up with one docker compose up next to Postgres, Redis and a worker. Built for events where CTFd starts to choke: two hundred challenges and up, hundreds of teams. The name comes from Reduta Ordona, a small redoubt meant to hold under an overwhelming assault. The code is not public yet, as the license is still to be decided before the first release.

• A precomputed scoreboard with live updates
• Bulk admin actions on challenges, with undo
• Challenge blocks with time windows and unlock rules
• A plugin protocol, including a reference King of the Hill hookup

Attack and defend: break into the box, then keep it.

A competition where breaking in is only half of it. Four vulnerable machines: a team gets in through the service, escalates to root and plants its beacon. Then the harder half starts, because points accrue for the time a box is held, while the entry vulnerability and the privilege escalation path come back on a schedule. No single patch closes a hill for good.

• Scoreboard and agent in the Python standard library alone
• Reports signed with HMAC, the key never enters the box players own
• Points for holding time, flags, first blood and service uptime
• Every hill is its own docker compose project with its own reset

Fourteen holes in one app, with a Star Wars story around them.

A deliberately vulnerable Flask web application written as a narrative: from the rebel base to the Emperor's dark archive. Fourteen flags ramped from level 0 to 5, each behind a different technique: information disclosure, XSS, SQL injection, broken access control, and a header the server trusts a little too much. Built as my Harvard CS50x final project, now used in student club classes and as a CTF challenge.

• Fourteen flags with categories and levels following the CTFd model
• No accounts and no login: every visitor gets their own run
• A separate bonus box with a privilege escalation path
• Meant for a local machine or a host cut off from the network