1Highlight: OpenAI launched GPT-6 Cyber; it says the model found 81 new high-severity Firefox bugs and built exploits for 948 Linux privilege bugs.
2Southeast Asia: AWS Semarak aims Amazon Quick at up to 445,000 Malaysian officers; Microsoft puts its next Malaysia cloud region in Johor Bahru.
3Also today: Nitori shows why data labels unlock AI analysis, a no-click fake AGENTS.md attack, and ChatGPT Images 2.6 (50% faster, GIFs).
Highlight of the day
AI security
81
new high-severity Firefox bugs OpenAI says GPT-6 Cyber found
Merged from 3 sources
★ Highlight of the day
OpenAI's GPT-6 Cyber finds 81 Firefox bugs and 948 Linux exploits
OpenAI launched GPT-6 Cyber, a model built to find and test software holes for security teams that are allowed to use it. OpenAI says it scored 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, a 51% relative gain. Access starts with approved Daybreak OpenAI's programme that gives approved security teams stronger AI tools to find and fix software bugs. Red customers only.
What we learned
Real systems: OpenAI researchers used it to develop working Exploit A way to use a software bug to make a system do something it shouldn't, like let a stranger in. for 948 Linux Privilege escalation A bug that lets someone gain higher computer access than they are supposed to have. bugs and to find 81 previously unknown high-severity Firefox bugs. Mozilla has begun shipping fixes. Cisco, Calif, CodeRabbit and others say they are using it for research and Code review Checking code someone (or an AI) wrote before it goes live, to catch bugs and risky changes..
Staying in bounds: OpenAI trained the model to stay inside the task a researcher sets. In one test, it avoided attacking out-of-Scope The exact target a security tester is allowed to attack; going outside it is off-limits. services in 99.93% of runs without a safety prompt, and 100% with the prompt.
Risk level: OpenAI rates the model Critical for cybersecurity under its Preparedness Framework OpenAI's own scale for how risky a model's skills are in areas such as cyber attacks., up from High for the previous Cyber model. It pairs the release with stronger misuse checks and a path for critical-infrastructure defenders.
Who can use it: GPT-6 Cyber starts on the API Application programming interface: a door that lets one piece of software ask another for data or actions. and Codex OpenAI's AI coding helper, which can write, fix and test software for you. for a select Daybreak Red group, then expands. Daybreak Blue access for related defender tools is also widening.
Why it matters: Security teams may find bugs faster, but the same Skill A saved, reusable set of instructions (and sometimes code) that an AI agent loads when a task needs it. can be misused. Only approved users get the strongest tools, and ordinary readers should treat viral 'leak' claims with care.
GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents · OpenAIRead ↗
OpenAI announces cybersecurity model GPT-6 Cyber; ExploitGym score up 51% · ITmedia AI+Read ↗
OpenAI's strongest cyber AI model GPT-6 Cyber; 81 new Firefox bugs · IThomeRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
More stories today
Southeast Asia
445,000
government officers AWS aims to reach with Amazon Quick
Merged from 3 sources
Southeast Asia
AWS aims Amazon Quick at 445,000 officers; Microsoft picks Johor
Two big cloud firms moved on Malaysia's AI plans on 5 November. AWS launched Semarak AI Malaysia to help the public sector, businesses and communities build with AI. Microsoft said its next Malaysia Cloud region A group of data centres in one area where a cloud company runs computers you can rent., Southeast Asia 3, will be in Johor Bahru.
Key learnings
AWS Semarak: three pillars cover government (AI Negara), business (AI Ekonomi) and communities (AI Rakyat). It builds on AWS's about RM29.2 billion (US$7.21 billion) Malaysia commitment.
Government tools: GovTech A government unit that builds digital tools for public services. Malaysia is launching GovHub with AWS's Kiro An AWS coding tool that helps developers write and change software with AI help. coding tool. A pilot starts with 500 developers in 2026, aiming for 4,000 by 2028. AWS is also investing to put Amazon Quick Amazon's workplace AI assistant that helps with research, summaries and data questions., its work assistant, in front of up to 445,000 government officers for tasks such as reading procurement data and policy research.
Small firms and schools: with GDEX's GD XCHANGE and VSTECS, eligible small businesses can get six months of free Quick and Kiro after an AI readiness check, aiming for 10,000 firms by 2028. More than 3,000 students and lecturers joined a polytechnic AI contest, AWS says.
Microsoft Johor: Southeast Asia 3 is part of Microsoft's about RM8.9 billion (US$2.2 billion) four-year Malaysia plan. Backup generators will use renewable Biofuel Fuel made from plants or waste instead of oil, used here in backup generators.. Microsoft says more than 734,000 people were trained under AIForMYFuture by October 2025, toward an 800,000 target by 2027. Digital Minister Gobind Singh Deo also urged that Google's RM61 billion plan create Malaysian tech, not only more computing.
Why it matters: Malaysia is stacking Data centre A large building full of computers that store data and run online services and AI, and that uses a lot of power and water., cloud regions and AI tools for civil servants and small firms at the same time. The officer and training numbers are company or programme targets, so delivery still needs watching.
AWS launches Semarak AI Malaysia, putting Amazon Quick tool to 445,000 government officers · TechNode GlobalRead ↗
Microsoft to open Southeast Asia 3 cloud region in Johor Bahru, per $2.2B Malaysia plan · TechNode GlobalRead ↗
Gobind: Malaysia must turn Google's US$15bil investment into homegrown AI innovation · Digital News AsiaRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
How companies use AI
2,600
main tables Nitori had to make AI-readable
How companies use AI
Japan's Nitori found AI analysis fails until people name their data
Furniture retailer Nitori moved sales data into Google BigQuery Google's cloud database made for analysing very large sets of business data. and ended up with about 2,600 main tables. Staff who already knew the data thrived; everyone else got lost. The fix was not a smarter chatbot. It was writing plain labels for what each column means.
Key learnings
The problem: with so many tables, only a few experts could still see the whole picture. Off-the-shelf chat tools failed because column names followed Nitori's own codes, such as DELVTO_CRP_KBN for 'delivery destination company type'.
The work: the IT team added meaning and Context window How much text an AI can keep in mind at once, like the size of its short-term memory. into table and column properties, linked them to an internal glossary, and used small AI agent An AI program that takes actions for you, like writing code, paying a bill or changing a menu, not just answering questions. for specific jobs instead of one giant agent.
What changed: people without SQL Skill A saved, reusable set of instructions (and sometimes code) that an AI agent loads when a task needs it. can ask for figures in everyday language. New managers can learn the data layout faster. Gemini can also help rewrite old data-cleaning jobs once the labels exist.
Lesson from Nitori's Kei Kobayashi: you do not need every table perfect. Labelling the top few most-used ones is enough to unlock most of the gain.
Why it matters: Buying an AI chat box does little if your spreadsheet names only make sense to three people. Clear labels are the boring step that makes AI useful for the rest of the team.
Nitori: how naming rules blocked conversational BigQuery analysis, and how labels fixed it · ITmedia EnterpriseRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
AI security
0 clicks
needed for a website to plant the fake instruction file
AI security
A fake AGENTS.md in Downloads can hijack coding agents with no click
Security firm Forever Security says it caught a new attack on real company computers. A website quietly drops a fake AGENTS.md A text file in a folder that tells AI coding agents how to behave there. Agents read it automatically. into the Downloads folder. Coding agents that trust that folder then rewrite their own settings and open the door to Malware Harmful software that steals data, spies on you or damages your computer..
Key learnings
How it works: tools such as Codex OpenAI's AI coding helper, which can write, fix and test software for you. and Anthropic Claude Code read AGENTS.md or CLAUDE.md for instructions. Browsers can also save files to Downloads without a click. Together, that means an agent can treat a planted Markdown file like a program.
What the fake file did: one real sample told Codex to edit ~/.codex/config.toml and install a bad npm package A shareable piece of JavaScript code that developers install into a project; fake ones can be malware. called function-color. A CLAUDE.md test rewrote .claude/settings.json so traffic went to an attacker's server about half the time.
How common: Forever Security says about half of its customer firms ran coding agents from Downloads, and about a quarter of agent sessions opened a browser.
Five free checks: do not run agents from Downloads; read unknown AGENTS.md or CLAUDE.md first; inspect config.toml and settings.json for strange servers; keep approval gates on; if you find a bad file, treat Token A small chunk of text, roughly part of a word, that AI models read and write. AI usage and pricing are counted in tokens. as leaked and reset them. Finding function-color installed is a red flag.
Why it matters: This is not a bug the vendors can simply Patch A small software update that fixes a bug or closes a security hole.. Keep agent work folders away from places the web can write to, and skim instruction files before you let an agent follow them.
Fake AGENTS.md in the Downloads folder can hijack coding agents without a click · ITmedia EnterpriseRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
AI products
50%
faster image generation vs Images 2.5, OpenAI says
Merged from 2 sources
AI products
ChatGPT Images 2.6 is 50% faster and can make GIFs
OpenAI released ChatGPT Images 2.6 on 4 November for every ChatGPT and Codex OpenAI's AI coding helper, which can write, fix and test software for you. plan, including free. It says generation is about 50% faster than Images 2.5, edits keep character details better, and GIF animation is rolling out.
Key learnings
New tricks: users can ask for looping GIFs or turn a sketch into a waving character. Saved reference photos can be named once, then reused in later chats with an @ mention, so you do not re-upload the same dog or product shot.
For developers: the API Application programming interface: a door that lets one piece of software ask another for data or actions. offers a faster Flare option and a higher-quality Sunburst option, plus SVG export and a complexity setting that controls how much the model plans before drawing.
Coverage: ITmedia and QbitAI both report the rollout. GIF support is still arriving in stages.
Why it matters: Faster edits and reusable references matter for small teams that make social posts and product shots. Treat GIFs as a convenience, not a reason to skip a human check before publishing.
AI security
