Solving government's most critical national challanges with sovereign AI and cyber-defense.
Global
Joined August 2026
- Tweets53
- Following31
- Followers171
- Likes25
5/ Governments face a growing wave of vulnerabilities while maintaining decades of legacy software, unmapped dependencies, and critical services they cannot switch off.
Defense must identify what is exposed, where attackers can go next, and how to stop them.
6/ Dream used Hero, our autonomous AI research system, to reconstruct the operation from attacker tools and logs.
The attackers went from an exposed firewall to an internal network in hours. Government defense must meet that pace.
Full research: dreamgroup.com/blog/chained-…
Check everything.
Dream’s autonomous cyber research platform, Hero, found an undocumented service that forwards authenticated session tokens to a cloud service without permission in a common industrial router made by ZBT.
This was just one of 15 security findings in a single firmware image, all found without human guidance.
In today's vulnerability discovery world, every government and critical infrastructure operator has off-the-shelf equipment that can carry new risks buried
Time to investigate it all.
Full research at dreamgroup.com/blog/who-cont….
This research was conducted on the ZBT router following significant findings in other ZBT routers conducted by VulnCheck.
Dream retweeted
Dream Co-Founder and CEO Shalev Hulio says rogue teams of two can do in minutes what used to take a 200-person state-sponsored team a year:
"AI gives capabilities of nations to everyone that has access to ChatGPT or Deepseek."
"Suddenly you are in a world that instead of having few threat groups like China or Russia or North Korea or Iran, every man with a computer and access to AI could be a big, big threat."
"Attackers are using AI, you have to build much stronger AI to be secure."
Some stats on #php2shell. In the 10 hours before it was publicly announced, across the @Cloudflare network we saw thousands of attempts to exploit it. That jumped to tens of thousands in the 45 minutes post announcement, hundreds of thousands within 2 hours, and millions now.
PHP powers WordPress, Wikipedia, and a huge share of the web.
PHP2Shell is especially challenging for governments. Finding every affected system means tracking PHP across ministries, municipalities, contractors, and internal tools, all across different owners and different update cycles.
The cost of exposure is massive as well. National adverseries are patient and tend to focus on lateral movements, which means a compromised website is just the first stop for something much bigger.
Depending on its access, that server could expose credentials, databases, or routes into internal systems.
Which server gets fixed first should depend on what an attacker could reach through it.
We’re already seeing early signs of AI-driven attacks exploiting PHP2Shell.
Dream is actively supporting our government customers respond, both by mapping exposure and attack paths for prioritization and then driving full remediation with Hero, our autonomous cyber researcher, by developing and validating tailored patches.
.@Shalevhulio on the state of the cyber threat today
Dream Co-Founder and CEO Shalev Hulio says rogue teams of two can do in minutes what used to take a 200-person state-sponsored team a year:
"AI gives capabilities of nations to everyone that has access to ChatGPT or Deepseek."
"Suddenly you are in a world that instead of having few threat groups like China or Russia or North Korea or Iran, every man with a computer and access to AI could be a big, big threat."
"Attackers are using AI, you have to build much stronger AI to be secure."
Come say hi to the Dream team at GTC this week. Will be on a sovereign data ontology panel, on demo lane and at booth 3019.
Berlin is an epic location for @NVIDIAGTC!
Follow along for announcements from Jensen's keynote tomorrow.
We’re releasing these research models on @huggingface so others can build on the work (but noting that they are separate from Dream’s production models).
We’ll also present the full paper at @emnlpmeeting 2026 in Budapest this year. The full research blog post is on are website at dreamgroup.com/blog/building…
The extra step was mid-training with a twist.
We turned 220,000 curated security documents into explanations, Q&A and analysis, then trained the model on that material before teaching it to follow instructions and chat.
Building AI for national cybersecurity requires more than powerful general-purpose models.
Governments need specialized AI that can operate securely and independently on infrastructure they control.
At Dream, we applied MiST, our cybersecurity mid-training method, to @nvidia Nemotron 3.5 Super VL, a 120B-parameter open model.
The results improved performance across all 15 cybersecurity knowledge benchmarks, saw average security score increase from 65.4 to 74.1 and outperformed other tested 120B-class open models on 12 of 15 benchmarks, all without impacting general capabilities.
Full technical deep dive, including how an AI agent and an experienced researcher helped accelerate training by 3.6 is on our research blog
dreamgroup.com/blog/mid-trai…
