Open to security roles — Berlin / remote in Germany

Prateek Pulastya

Security Engineer — AI & Application Security

I build and break the controls around LLM-powered systems. Most recently a two-tier prompt-injection filter that resolves the common path in 0.14 ms and is measured against four external benchmarks — including the ones where it does badly.

Based in
Berlin, Germany
Languages
English (C1+) · German (A2) · Spanish (A2)
Focus
AI/LLM Security · AppSec · Product Security
Certified
AWS Security – Specialty · Solutions Architect · CLLMSP
Experience
In cybersecurity since 2020 · independent research since 2025
Education
MSc Cybersecurity · MSc AI (in progress)
01

Selected work

Everything below links to source

Featured · Open source · Python

GuardRail-as-a-Service — prompt-injection detection for production LLM serving

A two-tier pipeline that filters prompts before they reach the model. Tier 1 is a deterministic Aho-Corasick + regex matcher that clears the common path in under 0.1 ms; Tier 2 is a DeBERTa-v3 ONNX classifier that only ever sees what Tier 1 lets through.

Python ONNX Aho-Corasick DeBERTa-v3 Docker Prometheus / Grafana Semgrep · Bandit · Safety in CI
The number I lead with is the weaker one

Tier 1's blocklist was tuned by inspecting the attacks it missed on my own corpus, so its 1.000 in-sample recall is partly memorisation. Split 60/40 and refitted on the training half only, it scores 0.7895 on held-out data — a 21-point overstatement if I quoted the in-sample figure. The repo says so in its own README, labels the contaminated row “not a result”, and stores every number as reproducible JSON under results/.

Measured performance

Corpus: 271 samples (190 attack / 81 benign) across 10 attack classes. Local Docker, CPU inference.
Evaluation Tier 1 recall FPR Reading
In-house corpus (in-sample) 1.0000.000 Fit, not generalisation
Held-out split, train-fitted rules 0.78950.000 Best estimate of real performance
Open-Prompt-Injection (source="data") 1.0000.006 External benchmark
BIPIA (source="data") 0.4800.017 Where it underperforms
NotInject (benign only, over-defense) 0.000 No over-blocking at Tier 1

Latency p50 / p95 / p99 — 0.14 ms / 26.21 ms / 32.11 ms. Tier 1 alone: p50 0.09 ms.

Engineering decision

Tier 2 ships monitor-only

On the NotInject over-defense benchmark the classifier blocks 40.4% of benign prompts against Tier 1's 0.0%, and the over-defense is not threshold-separable. So it scores and logs a TIER2-FLAG but does not block. The cost is a recall ceiling, stated explicitly rather than hidden; the blocking mode stays reachable via POST /validate?tier=2.

Engineering decision

Double scan, and fail open

Tier 1 scans normalised text first (keeping |, _, - so markers like <|im_start|> still match), then a de-leeted, separator-stripped variant — which is what catches 1gn0r3 prev10us 1nstruct10ns. If the Tier 2 model errors, the request is allowed: availability over coverage, deliberately.

Baseline comparison — read carefully

MetricGuardRailLlama-Guard-3-8B
Recall1.00000.1316
Precision1.00001.0000
Latency p500.16 ms1708 ms

This is not “7× better than LlamaGuard”. Llama-Guard-3 is a content-safety classifier whose S1–S13 taxonomy does not include prompt injection — it missed 165/190 attacks because it is not an injection detector. The defensible claim is that content-safety guardrails do not transfer to prompt-injection detection. The latency figure is local Ollama inference on a GTX 1650, not a hosted API round-trip, and the results JSON records that so it cannot be misquoted downstream.

Skills demonstrated
Adversarial ML evaluation Generalisation / held-out testing ONNX inference Latency-critical design Detection engineering Honest benchmarking

Project · AppSec · Detection-as-code

SecureFix

Six classic web vulnerabilities — SQLi, IDOR, SSRF, XSS, JWT-signature bypass, prototype pollution — each taken full circle: a working exploit, the fix, a hand-written Semgrep rule, and a CI gate that fails the build if it returns. The SAST gate goes 7 → 0 between the vulnerable and fixed branches. Both sides of the desk: finder's report and defender's PR.

TypeScriptSemgrepCI/CD gate

Live walkthrough →  ·  Source →

Skills demonstrated
Secure code review Detection-as-code CI security gates Exploit development OWASP Top 10

Project · AI security

ATSGuard

Stops indirect prompt injection — hidden white-on-white instructions in a résumé that hijack an LLM-based applicant tracking system. A four-layer pipeline reads styling, not just text, strips invisible Unicode and homoglyphs, and routes every file allow / flag / block. 0% false positives across 1,200 résumés; 98.8% in-sample recall — and 42.9% held-out, reported honestly.

PythonOWASP LLM01Prompt injection

View source →

Skills demonstrated
Indirect prompt-injection defence Unicode / homoglyph normalisation Adversarial evaluation Held-out reporting

Project · ML security

Explainable threat detection

Network intrusion detection over the CICIDS2017 benchmark across six attack classes, with SHAP attribution attached to every alert. Built for the actual SOC problem: not “is this an attack” but “which features made you say so”, so triage doesn't start from a black-box score.

PythonSHAPCICIDS2017

View source →

Skills demonstrated
Explainable AI (SHAP) Intrusion detection Feature attribution ML evaluation

Project · IoT · Top 10% extern

Secure MQTT pipeline

Hardened an unsecured hotel water-monitoring deployment through four independent layers — TLS 1.2/1.3 on 8883, mTLS client certs, HMAC-SHA256 message authentication, replay defense. STRIDE threat model first, then attack simulations before and after each layer to prove each one stops what the previous cannot. Added latency: under 1 ms.

MQTTmTLSSTRIDE

View source →

Skills demonstrated
STRIDE threat modelling TLS / mTLS Message authentication (HMAC) Attack simulation

Project · GRC

Security Compliance Academy

An open platform mapping technical controls to compliance frameworks — ISO 27001:2022, NIST CSF 2.0, SOC 2, GDPR. The differentiator is a cited, machine-readable control crosswalk: every mapping carries a version, a primary-source URL, and a confidence grade with a rationale, published as a fetchable /controls.json and validated in CI.

Next.jsISO 27001NIST CSF · SOC 2

Live site →  ·  Source →

Skills demonstrated
GRC Control-to-framework mapping ISO 27001 · SOC 2 · GDPR Data-schema design

Project · Applied AI

Mortgage document intelligence

Retrieval pipeline over 200+ page mortgage blobs: OCR (Tesseract, PaddleOCR), PyMuPDF parsing, LlamaIndex RAG tuned via chunk sizing and metadata filtering. Benchmarked OCR accuracy, retrieval quality and routing across Mistral and Phi-2 before recommending a deployment.

LlamaIndexRAGPyMuPDF

View source →

Skills demonstrated
RAG tuning OCR pipelines LLM evaluation Retrieval optimisation
02

Vulnerability research & disclosure

Outcomes reported as they landed

Independent research through private bug bounty and coordinated disclosure — #1 quarterly leaderboard (RO07 Ch13f) and Top 10 Bug Baron on Intigriti. Several findings below closed as duplicate or informative. They are listed with that outcome intact, because the methodology is the transferable part and a portfolio that only shows wins tells you nothing about calibration.

CVSS 9.9CWE-94 · via huntr

Sandbox escape in smolagents

Code-injection flaw in the agent execution sandbox permitting arbitrary command execution on the host — the boundary an agentic framework exists to enforce. Disclosed responsibly to the maintainers.

HighDetection bypass · via Intigriti

Path-traversal bypass in Aikido firewall-java

Payload construction that evaded the WAF's path-traversal detection while remaining a working traversal — a security control failing open against the exact class it advertises.

CriticalClosed: duplicate (valid)

CORS origin reflection → account takeover

A REST API reflected any attacker-supplied Origin verbatim while returning Access-Control-Allow-Credentials: true, with the session cookie scoped to the root domain and missing HttpOnly. One visit to a malicious page silently exfiltrated profile data, API keys and marketing contact databases as the logged-in user. Fix: server-side origin allowlist at the gateway, HttpOnly as defence in depth.

CriticalClosed: duplicate (confirmed valid)

Unauthenticated GraphQL user enumeration — 57M records

Introspection was disabled, but validation-error field enumeration exposed a Relay-style users connection accepting cursor pagination unauthenticated, with no rate limiting and no depth restriction. totalCount returned 57,316,627, confirmed live by a 209,852-account delta over eight days. Full export was reachable in roughly 32 hours at a conservative 10 req/s.

HighTriaged duplicate · P4

Unrestricted Google Maps API keys in page source

Two keys hardcoded into served HTML and a webpack bundle, both issued with no referrer, IP or API restriction — extractable with a single unauthenticated curl. A controlled 50-request load test returned 50/50 OK with zero throttling, establishing roughly $121/hour of billable abuse from one workstation. CWE-798; OWASP API4:2023.

InformativeClosed: theoretical exploitability

Dangling Route53 delegation → subdomain takeover

Certificate-transparency enumeration surfaced a staging subdomain with live NS delegation into an abandoned Route53 zone (SOA serial=1). Because Route53 returns deleted zones' nameservers to a shared pool, any AWS account could reclaim authority — and the root-scoped session cookie made that directly exploitable. Of 11 NXDOMAIN candidates checked via passive DNS and Wayback CDX, exactly one was still delegated.

03

Experience

Apr 2025 — PresentBerlin, Germany

Independent Security Researcher

Freelance / Self-employed

  • Security assessments across web applications, APIs, AI/ML libraries and cloud services through private bug bounty and coordinated disclosure programmes.
  • #1 Quarterly Leaderboard (RO07 Ch13f) and Top 10 Quarterly Leaderboard (Bug Baron) on Intigriti.
  • Disclosed a CVSS 9.9 (CWE-94) sandbox escape and a security-control detection bypass, using fuzzing (libFuzzer, CI Fuzz), static analysis (Semgrep) and custom Python tooling.
  • Built Python security automation for reconnaissance, vulnerability validation and log analysis while researching prompt injection, RAG security and AI evaluation methodology.
Dec 2025 — Feb 2026Remote

AI Systems & Document Intelligence Extern

Outamation

  • Built modular AI pipelines processing 200+ page mortgage blobs — OCR (Tesseract, PaddleOCR), PDF parsing (PyMuPDF) and RAG for extraction, classification and search.
  • Developed a LlamaIndex retrieval system tuned through chunk sizing and metadata filtering, evaluating Mistral and Phi-2.
  • Benchmarked OCR accuracy, retrieval quality and routing end-to-end; delivered a technical report on model trade-offs plus a demo UI.
Aug 2022 — Mar 2025Dublin, Ireland

Event Security Officer

Sword Security Group

  • Worked event security at 10,000+ capacity venues alongside the MSc in Cybersecurity, funding the degree and the relocation to Berlin.
  • Access control, crowd management and incident response across a 2.5-year tenure.
  • Coordinated multi-team operations during peak periods using standardised communication protocols.
Jun 2021 — Jul 2021Gurugram, India

Student Intern — Cyber Security

Gurugram Police Cyber Security Summer Internship

  • Supported live cybercrime investigations with Wireshark and Splunk, analysing network traffic and system logs to build evidentiary records.
  • Helped draft GDPR-aligned data classification and evidence-handling procedures, improving chain-of-custody integrity.
  • Delivered phishing and social-engineering awareness training for staff.
Feb 2021 — Jun 2021Delhi, India

Security Analyst (Penetration Tester)

Kratikal Tech. Pvt. Ltd.

  • Enterprise penetration testing for clients including GMR and Nykaa, identifying 70+ high-risk vulnerabilities across web applications, APIs and network infrastructure.
  • Delivered business-aligned remediation reports mapped to the OWASP Top 10 with prioritised fixes ahead of production deployment.
  • Supported ISO 27001 and NIST Cybersecurity Framework compliance assessments through structured control mapping.
Aug 2020 — Oct 2020India

Cyber Security Intern

Ethical Edufabrica Pvt. Ltd.

  • Manual penetration testing with Burp Suite, Nessus and Nmap across simulated enterprise environments.
  • Designed ransomware defence strategies and incident-response playbooks for controlled attack scenarios.
04

Technical scope

AI & LLM security

Prompt injection detectionRAG security AI red teamingOWASP LLM Top 10 NIST AI RMFGuardrail architecture MCP securityAdversarial evaluation

Application & API security

OWASP Top 10Threat modelling (STRIDE) Detection-as-code (Semgrep)Burp Suite Pro Secure code reviewSecure SDLC API securityCI/CD security gates

Offensive security

Penetration testingFuzzing (libFuzzer, CI Fuzz) Semgrep · BanditNmap · Nessus Vulnerability researchCTF

Cloud, detection & GRC

AWS (Security Specialty · SAA)Azure · OCI IAM least-privilegeMicrosoft Sentinel Defender for EndpointSplunk DockerPrometheus · Grafana ISO 27001 · NIST CSF · SOC 2 · GDPR

Engineering

PythonPowerShell ONNX RuntimeGitHub Actions PytestLlamaIndex
05

Credentials

Certificates held on file

Certifications

AWS Certified Security — Specialty

Amazon Web Services · Aug 2026 – Aug 2029 · ID …f67c44

Skills gained
Cloud security architecture IAM & data protection Threat detection & logging Incident response Infrastructure security
Verify →

AWS Certified Solutions Architect — Associate

Amazon Web Services · Aug 2026 – Aug 2029 · ID …18000c

Skills gained
Cloud architecture design High availability & resilience VPC networking Cost & performance optimisation
Verify →

Certified LLM Security Professional (CLLMSP)

Red Team Leaders · issued June 2026

Skills gained
OWASP LLM Top 10 Prompt injection MCP security Guardrail architecture NIST AI RMF
Verify →

Fortinet Certified Associate in Cybersecurity

Fortinet · Oct 2025 – Oct 2027 · ID 7813549201PP

Skills gained
Network security Firewall fundamentals Threat landscape
Verify →

Oracle Cloud Infrastructure 2025 Certified AI Foundations Associate

Oracle · issued August 2025 · ID 102394776OCI25AICFA

Skills gained
AI / ML fundamentals OCI AI services Generative AI basics

Certified Network Security Specialist (CNSS)

ICSI (International CyberSecurity Institute), UK · issued July 2020

Skills gained
Network security fundamentals Defensive controls

Recognition & education

Top Performer — Hydroficient IoT Cyber Defense Externship

Extern · April 2026 · awarded to the top 10% of participants

Completer — Outamation AI Document Insights Externship

Extern · February 2026

MSc Artificial Intelligence

Berlin School of Business & Innovation · Oct 2025 – Apr 2027 (in progress)

MSc Cyber Security

National College of Ireland · Jan 2022 – Feb 2023

Thesis: enhancing authentication and authorization in cloud services using blockchain

BTech Information Technology

Panipat Institute of Engineering and Technology · 2017 – 2021

06

Write-ups

Auto-synced from Medium every 6 hours

Loading write-ups…

07

Open source

Auto-synced from the GitHub API

Loading repositories…

08

Get in touch

Open to security engineering roles in Germany — AI/LLM security, application security, product security. Happy to walk through any result on this page, including the ones that did not go my way.