Every AI agent eventually asks for the same thing: real permissions on a real system. This is the gate I think it should have to pass.

agent ▸ request   write  prod/payments-db
        reason    "the migration test is failing, I can fix it"

guard ▸ what did it read?          retrieval provenance ............ checked   EmbedGuard
        what did it see?           secrets and PII on screen ....... masked    PerceptFence
        will the change hold?      validator veto + CI guardrails .. checked   TriArchitect · GenOps
        can we prove it later?     signed decision receipt ......... issued    RuntimeGuard-AI

guard ▸ ALLOW — with a receipt.

Building that gate is my job. I'm a Staff Security Infrastructure Engineer and the first security engineer at Parafin, which has funded $3B+ to 60,000+ small businesses and is joining Stripe (announced Sept 2026, close pending). Before that: Meta Reality Labs, Wayfair, JPMorgan Chase, American Express. Fifteen years, starting on the mainframe.

Each line in the gate is a question I have published on. Four peer-reviewed papers in 2026, two arXiv preprints, a merged probe in NVIDIA's LLM red-teaming tool, and a weekly newsletter, Securing the Agentic Stack.

Cross-Layer Detection and Provenance Attestation for Adversarial Embedding Attacks in RAG Systems · IJCESEN, Vol 12 No 1, pp 589–602 · 2026-02-05

An adversarial embedding looks right under cosine similarity and still steers the answer. EmbedGuard pairs embedding-space anomaly scoring with provenance attestation, so an attacker has to beat both layers at once. 94.7% detection on optimization-based attacks · 51 ms mean overhead · 500K embeddings · Docker-reproducible.

You share your screen with an assistant while debugging. Your terminal, .env file and CI logs get shared too. PerceptFence mediates what the model sees. The v0.4.0 test used 480 held-out developer-support screens, rendered in Chrome, degraded and read back by OCR, with the rules frozen before the test split ran:

The cost is measured too: 0.763 of task-relevant text survives on held-out screen types. Preprint, not peer reviewed, not counted among the four papers.

A Shared-State Multi-Agent Framework for Safe Java Code Migration · IJCISIM, Vol 18 No 18s, pp 739–751 · 2026-08-19

Agents drift when each one holds a private view of the codebase. TriArchitect gives them one typed shared state, the Typed Migration Graph, and a Validator-Veto Protocol: any validator can block an unsafe transformation before it lands.

A Governance-First Architecture for Embedding Generative AI into CI/CD Pipelines · JISEM, Vol 11(1s), pp 1518–1539 · 2026-02-15

Four pillars: contextualization, strategic guardrails, deployment assurance, governance audit trails. 55.7% lower median deployment cycle time (52.8 → 23.4 min, p < 0.001) · zero safety policy violations · 47.2% lower error-budget variance · 15,847 deployments across 127 microservices, 8 months, 3 organizations.

Scalable Tamper-Evident Accountability for High-Risk AI Systems Under the EU AI Act · JoCAAA, Vol 35 No 1, pp 1074–1090 · 2026-01-31

Article 14 asks for human oversight you can demonstrate. RuntimeGuard-AI uses Groth16 zero-knowledge proofs on bls12-381 to keep the audit trail tamper-evident inside a production latency budget. 2.3–4.1% latency overhead · 62 ms witness generation · 1,389 ms total proving at 50K constraints.

Durable Policy-Decision Receipts for AI Audit Evidence · arXiv:2608.17176 (cs.CR) · Aug 2026

The follow-up rebuilds RuntimeGuard-AI around the durability/latency trade-off. Every policy decision gets an Ed25519-signed receipt, and receipts chain into Merkle-epoch attestations. That receipt is the last line of the gate above. Preprint, not peer reviewed.

A gate is only as good as the attacks it has seen.

2025 ──────── PARAFIN ──────────────────────────────────────── present
 Dec          Staff Security Infrastructure Engineer
              ├─ First security engineer; built vulnerability management from zero
              │  (policy, asset tiering, priority matrix, SLA engine)
              ├─ Ran the SOC 2 Type II 2026 program end to end (12 workstreams, 86 controls)
              ├─ BC/DR from expired to exercised: live single-AZ production failover, zero customer impact
              ├─ Org-wide AWS baseline: tamper-proof CloudTrail (7-yr Object Lock), Security Hub, root MFA
              ├─ Led the company's first production TLS 1.3 migration
              ├─ OIDC + PKCE single sign-on for partner integrations
              └─ AI governance and MCP tool-access security

2024 ──────── META (REALITY LABS) ──────────────────────────── 2025 Dec
 Jun          Staff Software Engineer
              ├─ Tech lead, Reality Labs CI/CD modernization for Quest & Ray-Ban
              ├─ Build time −35% (P50), −42% (P95)
              ├─ GenAI migration tool: 95% manual effort reduction
              ├─ UBM metadata service: 450M+ requests/day, 99% fewer build-tracking failures
              ├─ Onboarding assistant: 55% oncall reduction, adopted by 3 teams
              └─ Mentored 6 senior engineers (3 promoted to Staff within 12 months)

2024 ──────── RAYMOND JAMES ────────────────────────────────── 2024 Jun
 Apr          Lead Software Development Engineer
              └─ Led a team of 8 on the client-onboarding document and e-signature platform

2022 ──────── WAYFAIR ──────────────────────────────────────── 2024 Mar
 Jan          Technical Architect
              ├─ DevSecOps platform adopted by all 15 business units
              ├─ Vulnerability management: 50M+ security checks/month
              ├─ Kubernetes policy-as-code engine (3% false positive rate)
              ├─ Unified OAuth 2.0: 125M+ API calls/week, 99.97% uptime
              └─ Mentored 20+ engineers (8 promoted to senior/staff)

2019 ──────── JPMORGAN CHASE ───────────────────────────────── 2021 Dec
 Mar          Senior Software Engineer (CIB)
              ├─ ESAP escrow portal (via Mphasis, then JPMorgan from Jul 2021)
              └─ Core platform and security lead

2011 ──────── AMERICAN EXPRESS (via TCS) ───────────────────── 2019 Mar
 Dec          Mainframe Programmer → Senior Software Developer
              ├─ International Insurance Services: monolith → microservices re-architecture
              └─ Bangalore → Montevideo → Phoenix

Security   Zero Trust · IAM · OPA/Kyverno policy-as-code · TLS/mTLS · Groth16 ZK proofs · MCP tool access
Cloud      AWS · GCP · Kubernetes · Terraform · Docker · ArgoCD · Datadog · PagerDuty
AI         RAG pipelines · LLM integration · agent red-teaming (garak) · MCP
Code       Java · Python · Go · TypeScript · Spring Boot · FastAPI · Kafka · PostgreSQL · Redis

28 technical articles + 13-issue newsletter · full list on ORCID. Also on Hashnode.

guard ▸ request from  you
        reason        "talk agent security, Zero Trust, or one of the papers"
guard ▸ ALLOW — no receipt needed.