Donato Capitella

Introducing spikee, an open-source toolkit for testing LLM applications against prompt injection attacks that can lead to exploitation—like data exfiltration, XSS, and resource exhaustion. Easily create custom datasets to match your specific use cases. spikee.ai

Agents are a new execution layer. Agent Commander is a promptware-powered C2. Red teaming is about to get weird. embracethered.com/blog/posts/202…

youtu.be/F8aOOp8VxFs Fast finetuning of LLMs like Gemma-3 on @AIatAMD Strix Halo (@FrameworkPuter Desktop) using Unsloth and distributed multi-node training.

@digitalix It took me ages fiddling with them, especially when the rack is full. It's infuriating!!!

the worst kinds of racks are the ones with these stupid little clips.

@wunderwuzzi23 Unfortunately it didn't make the cut 💔

@dcapitella Same. Looking forward to your Spikee talk also!

[un]prompted con next week! Hope to see you there. Your Agent Works For Me Now Lots of new demo exploits I haven't shared before and a deep-dive into delayed tool invocation No malware installed, no malicious binaries... JUST PROMPTS!

@Google it's good you're looking after your banned users xD

🔥 Took the Month of AI Bugs wreckage and turned it into a paper - AI Kill Chain 🧨 - Test cases and exploit chains (data exfil, rce, zombies!) - AgentHopper (a working AI virus for coding agents) 🦠 - SpAIware - Normalization of Deviance in AI zenodo.org/records/187692…

Google has totally dropped the ball on Antigravity.

Dual-GPU systems introduce real design considerations—from PCIe lanes to power and airflow. Join @dcapitella as he walks through a dual Radeon 9700 AI PRO setup and compares performance across single- and dual-GPU scenarios 👇 youtube.com/watch?v=dgyqBU…

Cline project was compromised. - Installer modified via a compromised token - For about 8 hours anyone who installed Cline also installed.... wait for it.... OpenClaw! - Can't make this up. Cline Security Advisory 👇 github.com/cline/cline/se…

In this video I introduce the new Strix Halo toolbox for Image and Video generation with ComfyUI, supporting SOTA open models like Qwen Image WAN 2.2 and the new LTX2: this is finally stable and performance has been improving. @FrameworkPuter @AIatAMD youtu.be/O57ideUzzTg

@digitalix Send one my way 😜

should?

Reviewing Claude Code output before pushing directly to prod

Evaluating multi-GPU LLM performance isn’t always straightforward. @dcapitella benchmarks vLLM on a dual Radeon 9700 AI PRO configuration and examines how tensor parallelism affects performance. Watch: youtube.com/watch?v=gvl3mo…

@SALTY_ALTY555 @FrameworkPuter Glad it helps!

@dcapitella @FrameworkPuter Oh my gosh this is going to save me soo much time.

youtu.be/nnB8a3OHS2E I built a 2-node Strix Halo/@FrameworkPuter Desktop setup using RDMA (RoCE/Intel E810) and vLLM tensor parallelism. This took a fair amount of digging into ROCm and RCCL to get low-latency multi-node inference working on gfx1151.

@Swissfold @FrameworkPuter I would get a switch instead! Probably 4 nodes would work well. I wish I could try that, but I am missing 2 more framework desktop boards and Intel E810 and the switch which can get incredibly expensive! :)

@dcapitella @FrameworkPuter Interesting ! How many nodes could be clustered in theory with this way ? I guess there is some similar cards with dual ports ?

@E_What_I_think @FrameworkPuter 🤣

@dcapitella @FrameworkPuter I keep trying to use my framework desktop as a desktop. Then you release stuff like this and I have to cluster both of my desktops.

I tried to make Jeff Geerling's @FrameworkPuter mini cluster faster. (new video)