.
contacto
cv // offensive ai-security

.

Offensive AI-security engineer — I build deterministic harnesses that prove and disprove agent exploits.

Offensive AI-security engineer — I build deterministic harnesses that prove and disprove agent exploits.

Offensive AI-security engineer — I build deterministic harnesses that prove and disprove agent exploits.

Chile · open to relocationdavid@dmzs.dev
GitHubgithub.com/raceksd-sourceORCID0009-0001-3950-2505LinkedIndavid-mellafe-zuvicarXivarxiv.org/abs/2606.05594
00
//

build

  • //deterministic LLM-agent evals
  • //oracle-as-sole-authority scoring
  • //TypeScript/Python tooling
  • //reproducible security datasets

break

  • //offensive security
  • //prompt-injection-to-tool-exploitation
  • //authorization-bypass research
  • //coordinated disclosure
01
//
[01]
→ public artifact

fail-closed network scope gate for AI agents/tools

Defends a real parse-differential scope-escape (gate-vs-resolver canonicalization divergence).

[02]
→ public artifact

deterministic answer-grounding oracle for LLM outputs

SHA-256 as sole authority — no LLM-judge, byte-identical, farm-resistant scoring.

[03]
→ public artifact

anti-fabrication guard

Flags hallucinated CVE ids in AI-generated security reports.

[04]
→ published research

sole-authored cs.CR paper

arXiv:2606.05594 v2; Zenodo DOIs 10.5281/zenodo.20501960 and 10.5281/zenodo.20501614. Finding: only 16/915 of Chile's Ley-21.663 critical-infra operators publish an RFC 9116 security.txt.

[05]
→ published tooling

published npm/TypeScript OSINT tool

~1k downloads; CI/CD via GitHub Actions OIDC. Related: latam-oiv-resolver alpha.

02
//

~6 vulnerabilities responsibly disclosed and submitted to MITRE (CNA-LR, IDs pending assignment); a plugin authorization-bypass (CWE-862) submitted to a major bug-bounty program; coordinated disclosures to Chile's national CSIRT / ANCI.

// methodology: fresh-code x dangerous-primitive, deterministic-oracle-confirmed before submission.
03
//
PythonTypeScript/Next.jsLLM & agent orchestrationdeterministic evaluation harnessesoffensive securityOSINT & threat-intel pipelinesRAG/Qdrant
04
//
[01]

— AlmaAI SpA / Reizan

Built deterministic eval harnesses, disclosure tooling, and reproducible security artifacts.

[02]

— KombuChile SpA

Prior operations leadership before moving into independent security research.

whatsapp