AI Sec Watch

ChatGPT has improved its ability to help with health questions through a new model called GPT-5.5 Instant, which better recognizes when urgent care is needed, explains uncertainty honestly, and provides clearer information. The improvements were measured using physician-led evaluations (HealthBench, a set of tests that assess health response quality) and real-world usage data, showing a 71% reduction in factuality issues over two months. GPT-5.5 Instant is available free to all ChatGPT users and performs similarly to OpenAI's most advanced models on health-related tasks.

AI-powered security operations centers (SOCs, where cybersecurity teams monitor and respond to threats) are automating many traditional analyst tasks, starting with alert triage and investigation. This shift will create new job roles such as security data engineers, AI security agent orchestrators, and AI model trainers, where humans will focus on preparing data, managing AI agent systems, and continuously updating AI models rather than doing routine alert monitoring.

AI systems challenge traditional cybersecurity because they behave unpredictably, unlike the deterministic (consistent and predictable) systems that security programs were designed around. Traditional security approaches focused on preventing attacks before systems go live, but AI agents make dynamic decisions and interact with external tools in ways developers can't fully predict, meaning security risks emerge at runtime (while systems are actively running) rather than being preventable beforehand. Additionally, AI-assisted development tools are accelerating code production, compressing the time security teams have to review and understand what enters production.

Researchers used OpenAI o3 Deep Research, an AI reasoning model, to re-analyze 376 previously unsolved rare genetic disease cases by connecting clinical data, genetic variants, and scientific literature into evidence-based explanations for human experts to review. After specialist evaluation and clinical confirmation, the AI-assisted workflow helped establish new diagnoses in 18 cases (4.8% additional diagnostic yield), with the model generating hypotheses rather than making medical decisions itself. This demonstrates how periodic AI-assisted reanalysis could help scale the process of solving rare disease cases as medical knowledge evolves.

Noam Shazeer, a senior Google engineer and co-lead of the Gemini AI models (Google's large language model system), has left the company to join OpenAI (the company behind ChatGPT). This departure highlights the competitive battle between tech companies to recruit top AI researchers and engineers.

Midjourney, known for its AI image generator, has unveiled The Midjourney Scanner, a hardware product that uses ultrasound technology (sound waves to create images of the body's interior) with a ring of sensors to capture full-body scans showing muscle, fat, bone, and organs. The company plans to build a spa in San Francisco where users could get these scans, which the CEO claims could match MRI (magnetic resonance imaging, a medical scanning technique) quality.

OpenAI is testing a new subscription service called 'ChatGPT for Science' that would provide specialized AI capabilities for scientific research, similar to how it previously created GPT-Rosalind (a specialized model built on GPT-5.5 architecture for life sciences research). The service would likely be restricted to verified research institutions and universities rather than being available to all users, and it is expected to be announced within the coming weeks.

Splunk Enterprise has a critical security flaw where a PostgreSQL sidecar service endpoint (a supporting service that handles database connections) doesn't require authentication (proof of identity), allowing an attacker without credentials to create or delete arbitrary files. This vulnerability is currently being exploited in real attacks in the wild.

Z.ai released GLM-5.2, a 753-billion parameter text-only open weights LLM (large language model, a type of AI trained on text) under an MIT license on June 16th, 2026. It features a 1-million token context window (the amount of text it can consider at once) and ranks as the top open weights model on independent benchmarks, though it uses significantly more output tokens per task than competing models. The model performs well on web development coding tasks but has shown mixed results in creative image generation tasks compared to its predecessor.