Research & news

Custom PHP2Shell Exploit Chained with Kernel LPE to Root Firewalls in the Wild

A router that phones home with your admin session tokens. That’s just one of fifteen findings in a single firmware image.

MiST adds a dedicated cybersecurity mid-training stage to compact language models. We are releasing MiST-8B and MiST-30B-A3B on Hugging Face.

How Dream builds deep security knowledge into a 120B open model on NVIDIA DGX B200, and what an AI agent and a researcher taught us about training it fast.

Dream uncovered an active DarkBlinders espionage campaign against Israel and the Kurdistan Region of Iraq, linking the activity to UNC5795, Dust Specter, and APT34 through GitHub-based command and control.

Fine-tuning small language models on up to 24 tasks at once left per-task accuracy intact and improved transfer to unseen tasks by as much as 10.8 points.

Hercules, Dream's on-premises AI vulnerability-research model, ranked #1 globally — built so defenders never send code to external frontier providers.

What gives state intelligence operations their edge isn't technical sophistication — it's a large specialized workforce. AI is now replacing that back office.

Open-weight models, mature agentic harnesses and cheap compute have put frontier-adjacent intrusion capability within reach of almost any operator.

Dream, the National Sovereign AI company, joined the NVIDIA-led Open Secure AI Alliance alongside Microsoft, IBM, Cisco and CrowdStrike.

Autonomous cyber operations have crossed from theory into practice. Amir Becker on machine-speed offense, attribution collapse, and the case for sovereign AI.

Reading a classification label straight from an LLM's logits in one forward pass runs 8-40x faster than generative classification, with no accuracy cost.