AI Radar Daily: GPT-5.5, Linux Root Vulnerability & DeepMind

Today is one of those days when the AI world once again splits very clearly into two camps: impressively useful and potentially extremely dangerous. Between autonomous attack simulations, a fresh Linux root vulnerability, and medical AI research, it becomes very clear how close progress and risk now sit to each other.

At the same time, there are also classic research topics that may generate fewer headlines but are important in the long run: probabilistic time-series models, latent-trained reasoning, and more efficient eigenvalue decompositions. In short: There’s a lot here for anyone who wants to understand AI not only as a product, but also as a technology.

🛡️ GPT-5.5 almost catches up to Claude in cybersecurity

The British AI Security Institute apparently tested GPT-5.5 the way a truly dangerous model should be tested: with autonomous network attack simulations. The result is remarkable: according to the reports, GPT-5.5 is only the second model that can solve a complete attack chain without human help. Claude Mythos still narrowly leads in the same tests — and it is not even freely available. GPT-5.5, on the other hand, is already inside ChatGPT and available via API. On the one hand, this is a technical milestone for reasoning and agent capabilities. On the other hand, it plainly shows how quickly powerful LLMs can become dangerous in cybersecurity scenarios. What used to be painstaking expert work is increasingly becoming a question of access control. For companies, this means security concepts must adapt to models that can do more than write text — they can orchestrate real attack chains. Source

🐧 Copy Fail: Linux root with 732 bytes of Python code

The new Linux security vulnerability with the charming name “Copy Fail” is anything but charming. According to heise, all major distributions have been affected since 2017, so this is a pretty broad issue. Especially unpleasant: the flaw can be exploited with just 732 bytes of Python to gain root privileges. That is the kind of “short proof of concept” that instantly makes admins take notice. This matters not only for classic server and desktop environments, but also for containers, DevOps stacks, and anything that relies on Linux kernel security. The practical lesson: kernel updates are not optional wellness programs, they are mandatory. And yes, “patch later” for root vulnerabilities is about as sensible as a lock on an open door. Source

📷 ChatGPT Images 2.0 hits a nerve in India

OpenAI is apparently getting a tailwind exactly where visual AI is currently especially ready to take off: in India. There, the new image features are being heavily used for avatars, personal images, and even cinematic-style portraits. That’s interesting because it shows that the success of GenAI products depends not only on model quality, but also on cultural fit, price sensitivity, and mobile usage behavior. If a feature takes off in one market, that does not automatically mean it will become the global standard. But it does give OpenAI an important signal: image generation is not just a toy for prompt nerds, but a mass-use case with very different regional tastes. For the market, this means localization and creative everyday usefulness are becoming more important than pure demo effects. Source

🧠 Probabilistic circuits for irregular time series

With “Probabilistic Circuits for Irregular Multivariate Time Series Forecasting”, we get a paper that sounds very academic at first glance, but addresses a real practical problem: irregularly sampled multivariate time series. These kinds of data are found constantly in medicine, industry, and monitoring — basically everywhere measurements do not arrive neatly and evenly every minute. The proposed architecture, “CircuITS,” tries to model uncertainty more cleanly while also avoiding contradictory predictions. That matters because classic forecasting models often look great until you unleash them on real-world data with gaps, outliers, and broken sampling intervals. For AI in practice, this is a good sign: research is moving closer to robust uncertainty models that do not just produce averages, but also credible probabilities. And that is exactly what you need when decisions are not supposed to rest on “probably fine.” Source

🔁 Latent-GRPO: Reinforcement learning in latent space

Latent-GRPO addresses a pretty exciting problem: reasoning models become more efficient when they do not have to spell out intermediate steps, but instead compress them internally into latent representations. That saves tokens and can significantly shorten chains of thought. The problem: reinforcement learning in this space is often unstable. That is exactly where the paper comes in, exploring how latent reasoning can be trained more stably with a form of group-relative policy optimization. For anyone who does not just use LLMs but wants to understand them, this is relevant because it targets the core of the next model generations: less visible “thinking,” more internal computational structure. In plain language: models could get smarter without reading every thought process back to us in full. Practically nice, scientifically exciting — and for anyone who likes debugging via thought logs, unfortunately a small step backward. Source

🩺 DeepMind is building an AI assistant for doctors

Google DeepMind is working on an “AI co-clinician” that is meant to support doctors in care delivery and even assess patients via video. In simulations, it looks promising, but it still falls short of experienced physicians. That is an important clarification, because the hype around AI in healthcare often moves faster than the evidence. Still, the work also shows where multimodal systems could become truly useful: pre-triage, documentation, history-taking support, and pattern recognition. At the same time, the article indirectly warns against misusing audio or voice-chat modes of ChatGPT for serious medical tasks. Good idea: AI as assistance. Bad idea: AI as a substitute for medical expertise. That is not an anti-AI statement, just common sense with a privacy bonus. Source

🧩 Batch-efficient eigenvalue decomposition for vision models

The paper “A Short Note on Batch-efficient Divide-and-Conquer Algorithm for EigenDecomposition” sounds like material for a very quiet conference audience, but it is relevant for many vision and deep-learning pipelines. EigenDecomposition is a bottleneck in numerous computer vision methods, especially when processing many small matrices in batches. The authors propose a more efficient variant that targets exactly that. This is not just numerical fine-tuning; in aggregate, it can save a lot of compute time, especially in models with geometric or structural components. For AI engineering, this is the kind of progress that rarely lands on front pages, but is worth real money in production systems. Or put differently: not every important AI news item has a giant language model in the title. Sometimes the matrix wins. Source

---

Don’t want to miss any news? Subscribe to the newsletter