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: Claude Code v2.1.287 turned mods on by default. A hub diagram shows a mod with unsandboxed access to files, secrets, prompts and network, approving tool calls before you are asked and spawning processes outside the sandbox. Mods install from any plugin marketplace, the org control is allowManagedModsOnly, and the community catalog lists 45 mods.

Claude Code Mods Are On by Default, and They Run Unsandboxed With Your Permissions

Claude Code 2.1.287 turns on mods by default: unsandboxed plugins that can read secrets and approve tool calls before you are asked. How to lock them down.

Oct 2 · 8 min

Data graphic: the mcp-atlassian MCP server flaw CVE-2026-77244 is rated CVSS 10.0 because its HTTP transport let anyone on the network act as the operator. "Fixed in 0.22.0" is struck through and corrected to "Run 0.23.1", since the SSE transport kept the same bypass until 19 Aug; versions before 0.22.0 are affected and no exploitation has been reported.

highAI Security Cve 2026 77244

MCP Atlassian Server Let Anyone on the Network Act as the Operator in Jira and Confluence

CVE-2026-77244 (CVSS 10.0): mcp-atlassian's HTTP transport ran unauthenticated requests with the operator's Atlassian token. Upgrade to 0.23.1.

Oct 2 · 7 min

Data graphic: MCP server steals OAuth secrets. CVSS 7.5 High 6.5 for the interactive provider ; affected mcp 1.9.1–1.29.1 and 2.0.0–2.1.1, fixed in 1.30.0 / 2.2.0, M2M providers must also pass issuer=. An attack path runs from the MCP client to a malicious MCP server to the auth server it picked, which receives the client_secret and PKCE verifier.

highAI Security Model Context Protocol

Official MCP Python SDK Let Malicious Servers Steal OAuth Secrets, and Upgrading Isn't Enough

A flaw in the official MCP Python SDK let a hostile MCP server redirect OAuth client secrets. Fixed in 1.30.0 and 2.2.0, but M2M providers also need issuer=.

Oct 2 · 9 min

Data graphic: 4-bit model file sizes against the memory budget for local LLMs on 16GB hardware. A 16GB Mac allows about 10.7GB of GPU memory by default, against roughly 15GB on a 16GB NVIDIA card, so Laguna XS 2.1 20.3GB and Qwen3.8-27B 16.8GB do not fit a 16GB card, while Devstral Small 2 24B 14.3GB , Gemma 4 26B A4B 14.2GB and gpt-oss-20b 12.8GB do. Practical context is 32K–64K for gpt-oss-20b, 8K–16K for Gemma 4 26B A4B and 4K–8K for Devstral Small 2 on the card, and 32K for Qwen3.5-9B ~5.6GB and 16K–32K for Gemma 4 12B 6.98GB on the Mac. The RTX 5060 Ti has 448GB/s of bandwidth to the RTX 4060 Ti's 288GB/s, and the verdict is to run gpt-oss-20b or Gemma 4 26B A4B on a 16GB card and Qwen3.5-9B on a 16GB Mac.

watchResearch Local Llm Inference

Best Local LLM for 16GB in 2026: RTX 5060 Ti, 4060 Ti and Mac mini M4 (Coding First)

gpt-oss-20b and Gemma 4 26B A4B lead on 16GB NVIDIA cards, while a 16GB Mac mini M4 is better served by Qwen3.5-9B. File sizes, context limits and VRAM math.

Sep 24 · 13 min

Data graphic: DGX Spark vs Mac Studio M5 Ultra vs Ryzen AI Max+ 395 for local LLMs. On gpt-oss-120b in the same llama.cpp test, DGX Spark prefills about 5x faster than Strix Halo 1,723 vs 340 tok/s , while decode is close 38.6 vs 34.1 tok/s ; the Mac Studio M5 Ultra has 1.2 TB/s of memory bandwidth against 273 and 256 GB/s, and price per GB of unified memory runs from about $27 Strix Halo boxes and $37 DGX Spark to $57 M5 Ultra base, 96GB and $42 M5 Ultra, 256GB .

watchResearch Nvidia Dgx Spark

DGX Spark vs Mac Studio M5 Ultra vs Ryzen AI Max+ 395: Which Box for Local LLMs?

Price per GB, memory bandwidth, prefill vs decode on gpt-oss-120b, clustering and software stacks: which unified-memory desktop fits your local LLM work.

Sep 24 · 14 min

watchResearch Vllm Vs Sglang

vLLM vs. SGLang in 2026: Benchmarking RadixAttention vs. PagedAttention for Local Agents & RAG

In agentic workflows, prefill latency and prefix caching matter more than raw tok/s. We benchmark SGLang RadixAttention vs vLLM PagedAttention and APC.

Sep 19 · 10 min

watchResearch Apple Silicon

Mac Studio (M4/M5 Ultra 128GB–256GB) vs. Dual RTX 5090 (64GB): The $6,000 Local LLM Workstation Shootout

Choosing a $6,000 local AI workstation: Apple Silicon Unified Memory (128GB-256GB) vs Dual RTX 5090 (64GB). Sizing 70B/236B MoE models, TTFT, and TCO.

Sep 19 · 9 min

watchResearch Local Llm Inference

The RTX 5090 (32GB) Local LLM Deployment Blueprint: NVFP4 vs. FP8, SGLang vs. vLLM, and Context Limits

Can the RTX 5090 run 70B models? We break down 32GB VRAM sizing, Blackwell NVFP4 vs FP8 benchmarks, KV cache math, and vLLM vs SGLang configurations.

Sep 19 · 8 min