Semantic prompt injection at the inference layer
How injection survives guardrails that operate above the model.
LEAD INVESTIGATOR
K. Rasel
Founder & Lead Investigator · LogicLeak
Most AI security audits are written by people who have never been on-call for the systems they are auditing. Mine are not.
— Operating principle, K. Rasel// PRACTITIONER PROFILE
three tracks of hands-on practice, converging on adversarial AI// 01 · VULNERABILITY RESEARCH
Software internals and exploit pathways
Hands-on dissection of vulnerability methodologies in enterprise software, with deep structural analysis of Foxit PDF and similar document-processing stacks. Approaches AI security the same way: not as a prompt-engineering puzzle, but as a software audit of a system that happens to include a statistical component.
Foxit PDF internals · enterprise document-processing vulnerability research// 02 · PRODUCTION INFRASTRUCTURE
Operating the systems most auditors only describe
Years of hands-on production infrastructure practice: server-side operations, remote environment migrations, and network security under live production constraints. Brings the perspective of someone who has been on-call for the kind of systems an audit might affect — not just someone who has tested them in lab conditions.
Production infrastructure operations · on-call experience// 03 · BARE-METAL AI
Local model deployment as a research instrument
Operates open-weight large language models on local hardware (DeepSeek, Llama via Ollama on Apple Silicon) at the bare-metal level. This is deliberate: hosted API endpoints abstract away the inference behavior where context bleed, semantic injection, and guardrail bypasses actually originate. Direct model access exposes failure modes that hosted endpoints hide from researchers.
M3 Max · Ollama · DeepSeek · Llama · open-weight research stack// AREAS OF ACTIVE INVESTIGATION
specific attack surfaces under current researchContext hemorrhage in long-running sessions
Token-window pollution as an attack and a cost vector.
Bare-metal vs API-mediated attack surface
Where hosted endpoints hide failure modes from auditors.
Multi-agent privilege inheritance
Cross-agent trust failure in tool-calling deployments.
Production infrastructure × AI deployment
Where AI-specific risks meet conventional infra weaknesses.
Document-processing pipelines as injection vectors
Markdown, PDF, and structured-document attack chains into RAG.
// PUBLISHED WORK
what we've written downFormal publications pending — first works appearing across the LogicLeak Research streams. See /research for upcoming pieces and the Field Reports published as they're sanitized for release.
SEE THE RESEARCH PIPELINE →// METHODOLOGY
Every engagement starts with a written scope, runs through reconnaissance and threat modeling specific to your stack, executes against your actual deployment under controlled conditions, and closes with a structured findings document and a remediation handoff. No part of this work is subcontracted to anonymous offshore teams. Sanitized findings may appear in LogicLeak Research quarterly — never with client names, never without your prior review. Read the full methodology at /research/methodology.
// CHANNELS
how to reach the investigator directlyEngagements start with a 30-minute scoping call. NDA before scope. 24h reply.
REQUEST AN AUDIT →