Adrian Sebastian Martin

Open to opportunities · Italy, remote anywhere

Adrian Sebastian
Martin

Breaks AI agents, measures the damage, publishes the record.

Penetration testing and AI agent security. AI security researcher, penetration tester or red team; junior roles and internships welcome.

See my workContact me
  • 5 public projects
  • 9 public records
  • September 6, 2026last verified

Every claim on this site links to a public record.

What I bring

Three strengths, each with its public record.

  1. AI agent security, proven in competition

    51st of 4186 teams and a silver medal in the Kaggle competition on multi-step tool attacks, hosted by OpenAI.

    AI agent security: multi-step tool attacks

  2. Published benchmarks, no LLM judge

    A defensive benchmark on agent scope discipline, published on Kaggle Benchmarks: 6 frontier models, 114 runs each.

    Agentic scope discipline benchmark

  3. Systems built and hardened end to end

    A production site, HTTPS only, with security headers anyone can verify; this site with no script or tracker and every link rechecked at each build.

    SigmaNex: site of the offline-first platform

51st of 4186 teams, silver medal.

Featured

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.

Work

Result first, record right under it.

  1. bit-flip: weight corruption in language models

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

    Experiments 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?

    Aug 2026 – today

  2. Agentic scope discipline benchmark

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

    A 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.

    Jun 2026 – Jun 2026

  3. LLM agent boundary evaluation

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

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

    Jun 2026 – Jun 2026

  4. SigmaNex: site of the offline-first platform

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

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

    May 2026 – today

All projects

About

I am a self-taught security researcher focused on penetration testing and on the security of AI agents: systems that call tools, act on networks and can be turned against their own rules.

I learn by building the target, attacking it and measuring the defence with evaluations that anyone can rerun. That method has produced public results: a silver medal, 51st of 4186 teams, in the Kaggle competition on multi-step tool attacks hosted by OpenAI; a defensive benchmark on agent scope discipline published on Kaggle Benchmarks and run on six frontier models; open research on weight corruption in language models, with code and data on Kaggle; and the site of SigmaNex, my own product, built and hardened end to end.

I am looking to join a team: open to remote roles anywhere as an AI security researcher, penetration tester or red teamer, including junior positions and internships.

Based in Italy

Skills

Each one links to a project that shows it.

Credentials

Each one links to its public verification page.

Path so far

  1. 2026

    Aug – today

    Experimental research on bit-flips in language-model weights

    Independent research

  2. 2026

    Jun – Sep 2026

    Featured Kaggle competition on AI agent security: silver medal

    Competition hosted by OpenAI on Kaggle

  3. 2026

    Jun – Jun 2026

    Defensive benchmark on agent scope discipline, published on Kaggle Benchmarks

    Independent research

  4. 2026

    May – today

    Founder and sole maintainer of SigmaNex, an offline-first intelligence and communication platform

    SigmaNex, own project

  5. 2019

    2019 – today

    Baker, working across every stage of bread-making

    Employment; study and research outside working hours

Open to remote roles anywhere: AI security research, penetration testing, red team, junior positions and internships. Every number on this site has a record behind it.

Write to me.

rootedlab@proton.me

Updated on