Dream

Solving government's most critical national challanges with sovereign AI and cyber-defense.

[Andrej Karpathy @karpathy]

Judging by my tl there is a growing gap in understanding of AI capability.

The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.

But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.

So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.

TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.

[Dream @DreamGroupAI]

See the full research report here https://t.co/jniLZDwOj0

[Dream @DreamGroupAI]

The conflict in the Gulf is also playing out through stolen credentials, compromised systems and intelligence gathering.

Our latest research, which leans heavily on our national cyberdefense capabilities, follows DarkBlinders, a campaign using fake email logins, shared files and meeting apps to gain access.

[Dream @DreamGroupAI]

At least 1 GB was stolen from a government cloud environment in Iraq’s Kurdistan Region. An individual linked to Israel’s security sector was also compromised.

The campaign remained active at publication and continues research initially posted by @Unit42_Intel .

[Sebastian Kurz @sebastiankurz]

Ever more drastic warnings about the potential consequences of “superintelligence” have sparked a lively debate about how to regulate it.

This week, the six leading AI companies gathered with President Trump at the White House to sign an accord—a sensible first step. Leading AI companies are committing to controls and independent audits of their most advanced models. Their first commitment includes preventing these models from accessing systems they are not supposed to enter.

But what responsible companies work to prevent is exactly what adversaries want to achieve.

You do not take off your seatbelt because other cars now have assisted driving. Your safety still depends on preparing for what others might do.

The same applies to nations. Governments cannot base their security on the assumption that everyone will act responsibly.

That is why we need AI not only regulated, but also under our sovereign control.

AI should keep advancing. Governments must ensure that our defenses advance with it. Voluntary commitments matter. But we cannot blindly rely on them to protect our people and our most critical assets.

https://t.co/sZq8Wz8ApM

[Shalev Hulio @Shalevhulio]

We are getting too comfortable with “the agent escaped.”

An agent escaping is a failure in development, architecture, or guardrails.

And this will get worse. Reports of agents probing government websites show how even a routine research task can become an attempted government intrusion.

Agents need limits that are enforced in software. This means permissions they can’t rewrite and mandatory supervision that can stop unauthorized actions before execution.

This need is one of the reasons @DreamGroupAI is Proud to be working with @NVIDIA to help solve this problem and build the security layer autonomous agents need.

If we can’t control it, we shouldn’t deploy it.

[Eytan Buchman @EytanBuchman]

Muse is the #1 downloaded app in the US this week.
Everyone's about to have an agent.

And if Australia's any lesson, it's all too likely that when you ask your agent to pay the electric bill, it hacks the power company by accident instead.

I live close enough to national critical infrastructure to know how thinly protected the long tail actually is.

We've seen the total number of capable threat actors balloon from open-weights, open-source harnesses, and even frontier model abuse.

But when you decouple intent from risk, it means a stratospheric increase in nation-state threats.

This is still just getting started.

[Sebastian Kurz @sebastiankurz]

Again. AI agents carry out a cyber attack. And the headlines are just the tip of the iceberg.

Just a short while ago, sophisticated cyber attacks took months of preparation, large teams and substantial resources.

Today, AI agents breach systems even without being instructed to. The incident in Australia highlights even more how quickly the threat is evolving.

Governments cannot rely on yesterday’s tools to protect against these capabilities. We need to put the best available AI to work defending our nations and their most critical infrastructure, from hospitals to energy and water supplies.

While this week has also indicated towards some sort of regulation, defenses have to go up in parallel.

https://t.co/Fbo3cZyCmP

[Dream @DreamGroupAI]

This one is for the @LangChain developers out there. One of our developers shares how and why some historical chat messages were changing when developing. Consider this a public service announcement. Full writeup here https://t.co/B4vwCR8AW0

[Dream @DreamGroupAI]

Governments don't rent their air force. They own it. AI should be no different.

Dream's Eran Hoffman on what sovereign AI actually requires, in a new interview with @IBTimesSG https://t.co/MHWD4NOJ6z

[Dream @DreamGroupAI]

As Dream's Karine shares with @sardesairajdeep on @IndiaToday, the road to protecting governments from AI attackers goes through deeper understanding of best practices and knowledge sharing with the right people.

[IndiaToday @IndiaToday]

AI Threat Alert: Hype Or Real?

'...Can we hit the pause button?' — Amid concerns over the Anthropic founder’s statement that AI is now generating new AI, @C_P_Gurnani says the development needs to be taken seriously. He suggests controlling the usage of AI and its use cases.

#ArtificialIntelligence #AISafety #FutureOfAI #AIThreat #NewsToday @sardesairajdeep

[Dream @DreamGroupAI]

Where does a small language model break when you keep piling on tasks?

We got to 24 without finding out. Per-task accuracy held, and transfer to unseen tasks kept climbing.

Full research here: https://t.co/BdAxpX9nK3

[Dream @DreamGroupAI]

How do you train the world's top-ranking autonomous cybersecurity research model and harness?

- Very specific training data set
- Rejection sampling
- Automated + manual tests
- and very, very smart people 💫

Elad Ben Zken, one of the AI researchers who worked on Dream model that ranked #1 globally on CyberGym last week shares more about the process, the model, and harness.

More technical information available here: https://t.co/kwRR97qLKr

[Shalev Hulio @Shalevhulio]

#1 in the world.

Dream just achieved the highest score ever recorded, 96.6% on CyberGym, UC Berkeley’s benchmark for autonomous AI vulnerability research.

I’m incredibly proud of this team.

We started @DreamGroupAI with the simple belief that in the age of AI, defenders must move faster than attackers.

Today is a strong proof point that we’re building exactly that, giving nations sovereign AI capabilities to defend themselves against the threats of the AI era.

And we’re just getting started.

[Sebastian Kurz @sebastiankurz]

We built @DreamGroupAI so nations can use AI to protect themselves and their most critical infrastructure  advanced cyber attacks.

It is deeply humbling to see that today, DREAM’s cyber-defense platform was ranked as the most powerful cyber defense AI in the world on the CyberGym benchmark.

This was not an easy task. It required the work of hundreds of highly qualified colleagues around the world.

But what matters is not the ranking.
What really matters is what it means for nations. That heating will stay on in the middle of winter, citizen data stays private and that hospitals keep running without disruption.

We will continue to work to provide nations the most capable national sovereign AI and cyber defense in the world.

[Dream @DreamGroupAI]

Governments need the best AI cyber-defense possible.
So that’s what we built.

Our cyber-defense agent and model just got ranked as the highest scorer on CyberGym, hitting 96.6% across 1,507 vulnerability research tasks.

This isn't just a research project. It’s already live and protecting governments and critical infrastructure in the real world.

More on how we built it:

https://t.co/kwRR97qdUT

[Dream @DreamGroupAI]

Governments operate with a constraint that cyber-adverseries do not. As @Shalevhulio shared with @ynetnews:
"...they will use any capability that gives them an advantage. Artificial intelligence will change the balance of power in cyber. Our responsibility is to make sure it gives defenders a greater advantage than it gives attackers..."

Full link here https://t.co/EdZyfPEW4W

[Dream @DreamGroupAI]

UAE. UK. Japan. Now Nigeria.

In one year, sovereignty over data, compute, and AI went from a regional story to a global one.

As @roblemos shared, the serious efforts all treat security as something verified continuously, not a one-time badge.

Dream CISO Lior Atar shares his thoughts on Nigera's National Sovereign Cloud with @DarkReading
https://t.co/g4F89RWvCt