1. Jun 2026 – Sep 2026Completed

    Role: Sole author

    Stack: Python, Jupyter

    AI agent security: multi-step tool attacks

    Multi-step attacks on a tool-using agent for the OpenAI-hosted Kaggle competition, with a working note that verifies its own 843 numbers.

    Result51st of 4186 teams, silver medal.

    In plain wordsI placed 51st of 4186 teams in a worldwide contest on tricking AI assistants into unsafe actions.

  2. Aug 2026 – todayActive

    Role: Author and maintainer

    Stack: Python, Jupyter, GGUF, Kaggle

    bit-flip: weight corruption in language models

    Reproducible experiments on bit-flips in language-model weights: random faults versus chosen bits, and whether alignment fails before capability.

    ResultSilent de-alignment not observed: de-alignment fraction 0 at every dose, on both arms (E5).

    In plain wordsExperiments on what happens to an AI model when tiny errors damage its memory: does it stop working, or does it keep working but stop being safe?

  3. Jun 2026 – Jun 2026Completed

    Role: Author and maintainer

    Stack: Python, Kaggle Benchmarks

    Agentic scope discipline benchmark

    Defensive benchmark: does an autonomous agent hold its authorization boundary when a hostile instruction arrives through a tool output?

    ResultPublished on Kaggle Benchmarks: deterministic oracle, no LLM judge, 6 models, 114 runs each.

    In plain wordsA public test that checks whether an AI assistant stays within what it was allowed to do, even when a web page or a file tells it otherwise.

  4. Jun 2026 – Jun 2026Completed

    Role: Author and maintainer

    Stack: Python, Kaggle Benchmarks

    LLM agent boundary evaluation

    Findings and research notes for the scope-discipline benchmark: seven frontier models measured with a deterministic oracle.

    Result7 frontier models measured, 63 runs each; offensive-action rate from 0.000 to 0.143.

    In plain wordsThe findings of the boundary test: seven leading AI models measured the same way, with numbers and notes anyone can re-check.

  5. May 2026 – todayActive

    Role: Founder and sole maintainer

    Stack: React, nginx, TLS

    SigmaNex: site of the offline-first platform

    Site of SigmaNex, an offline-first intelligence and communication platform: local AI, encrypted LoRa mesh, air-gapped operation. Front end and hardened server.

    ResultOnline at sigmanex.net, HTTPS only with HSTS, CSP and security headers anyone can verify.

    In plain wordsThe website of a product I built, a communication and AI system that works without internet, set up so that anyone can check its security.

Updated on