I work at the intersection of AI and offensive security — building threat models for LLM systems, researching adversarial ML, and helping organisations understand what breaks before attackers do. At SISA, I scope pen tests and write proposals for PCI DSS v4.0, ISO 27001, SOC 2, HIPAA, and DFIR engagements. Shipping proof of work before graduation.
→ Public AI security research write-up from SISA (GitHub or blog post)
→ 4 LinkedIn posts on LLM security — adversarial ML depth
→ Reading: FGSM · membership inference · model inversion papers
→ Portfolio site (no overthinking the design)
→ Reaching out to people 5–10 yrs ahead in AI security
| Repo | What it is | Stack |
|---|---|---|
| proposal-engine | 5-agent pre-sales pipeline — discovery → scoping → pricing → adversarial critique → proposal. 6 BUs, 22 services. Live on Streamlit Cloud. | Python · Groq · Llama 3.3 · Streamlit |
| presales-automation | Full-stack DFIR pre-sales tool — AI-generated DOCX proposals, opportunity pipeline, Claude API with prompt caching, exec dashboard | React · Express · SQLite · Claude API |
| claude-code-best-practices | Production patterns for Claude Code — CLAUDE.md setup, security rules, token efficiency, JARVIS AI dashboard | Node.js · TypeScript · Anthropic SDK |
| portfolio | Personal portfolio — dark web3 Next.js 15, multi-model AI chat (Groq + OpenRouter fallback), Vercel Edge Runtime | Next.js · TypeScript · Vercel AI SDK |
AI Security → Adversarial ML · LLM threat modelling · AI zero-day research · red-teaming AI systems
Security Testing → Web app · Network · API · Cloud pen testing · OWASP ZAP · VA&M
DFIR → Incident response · Digital forensics · Threat hunting · Tabletop exercises
Compliance → PCI DSS v4.0 · ISO/IEC 27001 · SOC 2 · HIPAA
Pre-Sales → Discovery · Scoping · Proposal generation · Pipeline automation
Languages
AI / Security
Web / Infra
Daily AI security insight — updated by JARVIS.
| Date | Insight |
|---|---|
| 2026-08-29 | LLM hallucination in security context: models confidently output fake CVE IDs — always verify |
| 2026-08-28 | LLM hallucination in security context: models confidently output fake CVE IDs — always verify |
| 2026-08-22 | LLM memorization scales with model size — GPT-4 class models memorize more training verbatim |
| 2026-08-08 | Claude's Constitutional AI: model critiques and revises own outputs — red-team the constitution itself |
| 2026-08-07 | Supply chain risk: fine-tuning on poisoned HuggingFace datasets is underexplored attack surface |
| 2026-08-02 | Adversarial suffixes (Zou et al., 2023) transfer across GPT-3.5/Claude — alignment is brittle |
| 2026-07-18 | Reward hacking: RL agent exploits reward model bugs rather than learning intended behavior |
| 2026-07-17 | Zero-shot transfer of adversarial examples across architectures — architecture diversity ≠ safety |
| 2026-07-14 | Multi-agent coordination attacks: one compromised agent poisons shared context window |
| 2026-07-11 | LLM-as-judge manipulation: adversarially crafted outputs fool automated evaluation pipelines |
| 2026-07-05 | Tool-use abuse: LLM agents with shell access are lateral movement vectors if prompt-injectable |
| 2026-07-04 | RAG poisoning: attacker controls a web page the retriever indexes → arbitrary instructions injected |
| 2026-07-03 | Adversarial suffixes (Zou et al., 2023) transfer across GPT-3.5/Claude — alignment is brittle |
| 2026-06-27 | Differential privacy in LLM training: ε=8 provides weak protection, ε=1 kills utility |
| 2026-06-26 | Differential privacy in LLM training: ε=8 provides weak protection, ε=1 kills utility |
| 2026-06-21 | Synthetic data laundering: train on AI-generated data to remove data lineage — legal grey area |
| 2026-06-20 | Agentic AI risk: irreversible real-world actions need hard guardrails before execution |
| 2026-06-15 | Embedding inversion: from output embeddings, reconstruct input text with ~75% token accuracy |
| 2026-06-10 | Gradient-based attacks (PGD, FGSM) on vision-language models can bypass safety classifiers |
| 2026-06-09 | Claude's Constitutional AI: model critiques and revises own outputs — red-team the constitution itself |
| 2026-06-08 | Supply chain risk: fine-tuning on poisoned HuggingFace datasets is underexplored attack surface |
| 2026-06-04 | RAG poisoning: attacker controls a web page the retriever indexes → arbitrary instructions injected |
| 2026-06-01 | Model inversion attacks can reconstruct training data — PII in fine-tune sets is a real threat |
| 2026-06-02 | Context window poisoning: long-context models are more susceptible to late-turn instruction hijack |
The repos are what count. Building proof of work before graduation.