Tech • AI • Robotics • Game
GPT-6 cyberattack, new video AI, autonomous agent and robots
9.5/10TL;DR
OpenAI, Google, Microsoft and Perplexity are advancing AI systems that can hack software, generate and edit video, alter live speech and automate office work, while robotics trials in China and South Korea show how quickly machines are moving into real-world jobs.
KEY POINTS
OpenAI unveils a specialized cyber model
GPT-6 Cyber is designed to find, exploit and help fix software vulnerabilities. It is not a general-purpose assistant but a security-focused model retrained with reinforcement learning to succeed at offensive cyber tasks, then produce remediation advice. Access is restricted to approved clients rather than offered broadly.
Performance jumps beyond guarded public models
On benchmark tests, the model reportedly approaches success rates near 50% on exploit tasks, far above a standard public version of GPT-6 Astra with guardrails enabled, which scores effectively at 0% in the same setting. The gap suggests this is not merely an unlocked mainstream model but a system specifically optimized for vulnerability research and exploitation.
Firefox and Linux findings raise the stakes
In tests on Firefox, the model identified 81 previously unknown vulnerabilities judged serious by Mozilla, plus 30 known flaws it reassessed as more dangerous than previously thought. Against the Linux kernel, it reproduced 1,724 vulnerabilities, including 788 unknown ones, and generated 948 privilege-escalation exploits that could grant administrator rights from a standard account.
Cheap cyber capability intensifies access concerns
The Linux effort reportedly consumed about $1.2 million in compute, which works out to a little over $1,000 per exploit found. That cost is low enough to sharpen fears that extremely advanced offensive cyber capability is becoming industrialized. OpenAI says it shared reports and patches with maintainers, but the episode also highlights how a private company now decides who gets access to a tool that can both secure and destabilize critical systems.
Google expands low-cost AI video production
Google released Gemini Omni 1.2 Flash, a new video model that can generate or transform clips from text, images, audio or existing video. It can output HD video with sound and extend clip length to 15 seconds, up from 10 seconds, a change with clear implications for online advertising and product marketing.
Video generation gets cheaper and more flexible
The model is priced at roughly $0.07 to $0.08 per second, putting a 15-second clip at around $1 depending on complexity. It is available through Google AI Studio and integrated into Google Ads and Google Vids, where users can create rough drafts at low quality before rerendering in 1080p. Google Vids also offers three free generations per month.
Microsoft patent targets live speech rewriting
A newly noticed Microsoft patent describes a system that could alter a live video stream with only about 10 seconds of buffer. The technology would not simply dub or translate speech: it could rewrite wording, smooth hesitations, modify tone and synchronize facial movements and gestures so the edited version appears natural in real time.
Personalized streams could reshape trust online
The patent also outlines the possibility of tailoring output to different viewers, including language and presentation changes. In its most benign form, that means smoother translation and cleaner delivery. In a darker scenario, it could produce individualized versions of the same live appearance, raising fresh concerns about authenticity, manipulation and the future reliability of audiovisual evidence.
Perplexity enters the race for persistent AI agents
Perplexity launched Teammates, persistent AI agents available first to Max subscribers at about €170 a month. Unlike one-off assistants, these agents can stay connected to tools such as Slack, GitHub and email, operate continuously in the cloud, and carry out recurring tasks like support, documentation, sales operations or research updates.
Multi-agent automation moves toward digital coworkers
The new service is model-agnostic and built on Perplexity Computer, giving agents their own browser sessions, memory and workspaces. A single agent can also call other agents, creating layered workflows that resemble departments rather than chatbots. That structure reflects a broader 2026 shift toward autonomous, persistent AI workers marketed as low-cost substitutes for routine knowledge work.
Robots are being tested in public and industrial settings
In Shanghai, AgiBot ran a nearly 23-hour live retail trial with six robots in a mock convenience store watched by roughly 2 million people. The machines stocked shelves, served drinks and handled customer disruptions; four were judged good enough to keep, while two failed due to slow performance or inability to handle edge cases. In South Korea, 20 humanoid robots from Robros are being cleared for tightly supervised factory pilots alongside humans in four industrial settings, testing whether barriers between workers and robots can safely come down.
CONCLUSION
The latest wave of AI is no longer limited to chat and image generation: it now reaches cybersecurity, live communication, office work and physical labor. The central question is no longer whether these systems are capable, but who controls them, how they are deployed and how fast institutions can adapt.
Ask a question