Adversarial testing, low-level optimization, and quantitative research.

Cross-model adversarial evaluation of four production LLMs. 107 vectors per model across the ten OWASP LLM Top 10 categories, canary-based deterministic detection, every run completed with zero errors. gemini-3.1-flash-lite, gpt-oss-20b, gpt-oss-120b and qwen3.8-27b were bypassed on 15, 10, 5 and 5 vectors respectively.

The 51 vectors covering jailbreaks, encoding bypass, indirect injection and system-prompt extraction produced zero bypasses on any model. All 35 observed failures fell in tool use, output handling, PII disclosure and confabulation — including one vector that failed on all four, where reformatting a config file as JSON reproduced its credentials verbatim.

These models refuse the attacks they are trained hardest to refuse. The exposure is downstream of the refusal — in what they emit, what they do with tools, and what they repeat back.

_{Single pass at temperature 0, an observation not a rate · synthetic harness prompt · suite now holds 300 vectors, additions unmeasured · raw reports and per-finding transcripts published in full}

Full evaluation →  ·  Raw data →

Find where an LLM's guardrails crack — deterministically, not by judge-model opinion.

A fresh random canary is planted in the system prompt at scan time, so a finding is a string match: reproducible, one API call per vector, no second model grading the first. 12 vectors run as real multi-turn conversations, because a guardrail that holds against one message often erodes across five.

lrtf scan gpt-4o --tui               # live view as each vector lands
lrtf scan gpt-4o --system mine.txt   # test your own prompt, not a toy one
lrtf compare gpt-4o claude-sonnet-4-5 ollama/llama3
lrtf diff base.json current.json     # did your fix actually work?

A backtester that tells you when your strategy adds nothing.

Signals shift one bar — no lookahead. Cost and slippage are charged against turnover. Every run reports alpha and information ratio against buy-and-hold, so a strategy with no edge says so out loud.

Walk-forward validation, ensemble blending, target-vol and fractional-Kelly sizing, Black-Litterman allocation over shrinkage and PCA-factor covariance, Fama-French overlay, bootstrap and Jobson-Korkie significance tests. Pure numpy/pandas.

qfcli --backtest AAPL --walk-forward --ensemble rank --grid "fast=10,20;slow=40,60"
qfcli --portfolio AAPL MSFT NVDA --bl --view NVDA=0.18 --ff

See what's listening on a network. No dependencies, no nmap underneath.

Native parallel TCP scanner in C++17. Non-blocking connect() awaited through poll(), from a bounded worker pool with deterministic concurrency — scan the same range twice, get the same behaviour twice. Banner grabs on open ports, JSON output, input validated before a single packet is sent.

Verified by a CTest suite that spins up real listening sockets, with ASan/UBSan in CI on macOS and Linux.

zapscan -p 1-1024 -c 256 scanme.nmap.org
zapscan -j -o report.json -p 22,80,443 10.0.0.0/24

Languages  C++ · Go · Python · x86_64 Assembly

LLM security  prompt injection · jailbreak and encoding-bypass vectors · system-prompt extraction · multi-turn attack chains · canary-based deterministic detection · OWASP LLM Top 10

Agentic AI  memory poisoning (LangChain / ChromaDB / Mem0) · RAG and retrieval security · tool-use and excessive-agency attacks

Offensive security  network recon and port scanning · dependency vulnerability analysis and CVE identification · low-level socket programming

Quant  no-lookahead backtesting · walk-forward validation · Black-Litterman · shrinkage and PCA-factor covariance · vol-managed sizing

• AI red teaming — tooling against the OWASP LLM Top 10, with detection that is deterministic rather than model-judged
• Agentic AI security — mapping memory-poisoning and agent-manipulation surfaces, and building runtime defense against them
• Invited maintainer — OWASP/cve-lite-cli