AI security researcher

Owen Park

Articles: 62
Desks: 2

Covers AI application security, model risk, and agentic system abuse.

ai-securityresearchapplication-security

Coverage by Owen Park

Data graphic: GLM-5.3-Flash's FP8 weights take 328 GB for 320B parameters 1.025 bytes per parameter . A bar chart of effective bytes per parameter shows FP8 and INT8 at 1.0 to 1.05 and the 4-bit formats lower: INT4 GPTQ/AWQ 0.52 to 0.60, NVFP4 0.56, MXFP4 and EXL3 0.53. Worked examples: 328 GB fits four 128 GB DGX Sparks 512 GB , Qwen3.8-Flash in NVFP4 at about 135 GB is 7 GB over one Spark, and a projected 4-bit GLM-5.3-Flash of about 170 GB is an estimate, not a file size.

NVFP4 vs FP8 vs MXFP4 vs EXL3: The Quantization Format Guide for Local LLMs

The same 320B model is 328 GB in FP8 and near 135 GB at 4 bits. Bytes-per-parameter math, per-format tradeoffs, and how to size a box that actually fits.

Sep 12 · 13 min

stableResearch Deepseek V4 1 Flash

Run DeepSeek V4.1 Flash on Three DGX Sparks

A stability-first SGLang deployment serves DeepSeek V4.1 Flash across three NVIDIA DGX Sparks: 256k context limit, 750k-token KV pool, ~38 tok/s single-stream decode.

Sep 12 · 3 min

Infographic mapping one to four NVIDIA DGX Sparks to the recommended local model at each node count: Qwen3.8 Flash Next, DeepSeek V4 Flash, GLM-5.3 Flash, DeepSeek V4.1 Flash.

stableAI Security Nvidia Dgx Spark

Best Local AI Models to Run on Your DGX Sparks

From one Spark to four: the open-weight model to serve at each node count, with measured decode, prefill and context figures from Mia's AI Lab recipes.

Sep 12 · 10 min

Data-poster diagramming an AI agent control plane: an agent routed through interception, workload identity, a policy engine and guard models to a credential broker, an MCP server and a signed audit ledger.

elevatedAI Security Agent Control Plane

The Agent Control Plane: Engineering Runtime Authorization for MCP and AI Tool Calls

Prompt injection resists a model-layer fix. Inside the agent control plane: workload identity, brokered credentials and blast-radius caps at the tool call.

Sep 4 · 12 min

ThreatFrontier poster showing Claude Code internals as a local agent runtime with tools permissions sessions and SDK

watchAI Security Claude Code

How Claude Code Works Internally: From CLI Startup to Agentic Tool Execution

A practical systems-level explanation of Claude Code as a local agent runtime, covering startup, authentication, tools, permissions, sessions, MCP, plugins, and the Agent SDK.

May 20 · 13 min

ThreatFrontier poster showing Langflow CVE-2026-33017 public workflow build endpoint remote code execution

criticalAI Security AI Gateway Security

Critical Langflow Vulnerability Exposes AI Workflows and API Keys to Active Exploitation

CVE-2026-33017 exposes Langflow AI workflow builders to unauthenticated remote code execution through the public flow build endpoint.

May 18 · 6 min

Protocol schematic poster for a 128GB Apple Silicon local AI stack led by Qwen3.6

watchAI Security Local AI

Best Open-Weight AI Models for a 128GB Apple Silicon Mac in 2026: Qwen3.6 Leads the Pack

Qwen3.6-35B-A3B leads 2026 open-weight AI picks for 128GB Apple Silicon Macs, balancing coding, reasoning, long context and local speed for developers.

May 17 · 6 min

Editorial poster showing a 48GB Apple Silicon local AI model stack for open-weight coding and RAG models

watchAI Security Local AI

Best Open-Weight AI Models for a 48GB Apple Silicon Mac in 2026

A 48GB Apple Silicon Mac is a strong local AI machine for quantized 24B-32B open-weight models, led by Qwen3-Coder-30B-A3B for coding and Gemma 4 for reasoning and RAG.

May 17 · 7 min