948 AI-found Linux root bugs, Tencent's Hy4 sees screens, ChatGPT makes GIFs
OpenAI says one run of an AI pipeline turned up 948 distinct ways to become root on Linux, at about $1,279 each. Tencent shipped Hy4, its first model that can read images and operate a computer, at budget prices. And ChatGPT can now make animated GIFs. The thread: finding bugs, reading screens and making visuals all got cheaper this week, and the work moves to whoever turns that into a dependable service.
Key takeaways
- OpenAI says an unreleased cyber model produced 948 unique Linux root exploits in one mid-October run for about $1.2M, but only 61 were patched upstream ten days later, so patching is now the bottleneck.
- Tencent's Hy4 adds vision and computer use at about $0.90 per million input tokens, but its benchmarks are self-reported and the open weights are not out yet.
- ChatGPT Images 2.6 brings GIFs, saved reference images and a faster API model, which lowers the cost of short animated and vector content for small businesses.
OpenAI says one AI pipeline run produced 948 unique Linux root exploits for about $1.2 million
On November 4 OpenAI described a mid-October run in which an unreleased cyber model found, exploited and drafted patches for 948 distinct Linux privilege-escalation bugs, at an API cost of about $1,279 each.
OpenAI published a research post on November 4 describing what happened when it pointed an upcoming, still-unnamed cyber model at the Linux kernel. In a single mid-October run, the company says its automated pipeline produced working exploits for 948 unique local privilege-escalation bugs. These are flaws that let someone who already has a foothold on a machine, such as an ordinary user account or a compromised app, take full 'root' control of it. OpenAI says the pipeline also wrote a candidate patch for each one. The work sits under Patch the Planet, a programme OpenAI introduced in June with Trail of Bits to get AI-found bugs fixed.
The reported funnel is detailed. The system scoped 5,034 targets covering about 2.8 million lines, roughly 7% of the kernel. Within 12 hours it produced 2,308 deduplicated candidate bugs; within 30 hours it had reproduced 1,724 of them. That led to 291 working information-leak demonstrations and 1,169 working root exploits, which OpenAI groups into 948 unique bugs by counting distinct patches. The run cost $1,213,430 in API credits, about $1,279 per unique bug, a figure OpenAI calls an upper bound. A similar run in late July with GPT-5.6 Sol found 166 unique bugs for $513,690.
OpenAI also says the agents used at least ten exploitation techniques that it and outside kernel experts considered new, based on a sample of 166 exploits, and that it is withholding the technical details for safety. Tests ran in isolated virtual machines with standard kernel hardening switched on. Interesting Engineering reports an illustrative comparison using past Pwn2Own and TyphoonPWN payouts that values the bugs at $18.9 million to $66.3 million, 16 to 55 times the pipeline's cost. That is a thought experiment, not money anyone was paid, but it shows how sharply the price of finding serious bugs is falling.
The caveats matter. The 948 is not a count of fixes in the kernel. By the end of a ten-day scan and analysis period, Unite.AI reports, 61 bugs had been patched upstream, 43 had patches in staging, and 299 had been reported or publicly discussed; the rest are to be shared gradually through partners. OpenAI says humans review each exploit, patch and severity rating before reporting, and admits that in an early run 20 of 200 'successful' exploits actually reached root through a different, unintended bug. All figures come from OpenAI and have not been independently audited.
The practical message for anyone running Linux is that the gap between a bug being found and a patch reaching your servers is now the weak point. If one company can find hundreds of root bugs in days, attackers with similar tools will find some too, and kernel maintainers will face a flood of reports. Expect more kernel updates, more security advisories and more pressure to reboot machines regularly. Small firms that run Linux servers, routers, point-of-sale terminals or cloud instances without a patching routine are the most exposed, and the least likely to notice.
Why it mattersFinding serious operating-system bugs is becoming cheap and fast, so the risk shifts to how quickly fixes are shipped and installed, and most small organisations have no routine for that.
👀 What to watch
Whether the Linux kernel's patch volume and security advisories jump in the coming weeks, when OpenAI names or releases the cyber model, and who else publishes similar runs.
3 opportunities from this story
Monthly Linux patching retainer for small firms
Offer small businesses a fixed monthly service that keeps their Linux servers and cloud instances on supported, patched kernels, with planned reboots and a short monthly report. Many firms run a few servers set up years ago by someone who has since left, and nobody checks kernel versions. OpenAI's numbers give you a concrete reason to start the conversation: more kernel bugs are being found faster, so patch frequency matters more. Sell reliability and a paper trail, not a promise that they will never be hacked.
- Best for
- Freelance sysadmins, DevOps contractors and small IT support firms
- First step this week
- List every Linux machine you already manage with its kernel version and last reboot date, and turn that into a one-page sample report.
Open the full playbook
Launch steps
- Inventory each client server: distribution, kernel version, uptime, who has access
- Turn on automatic security updates or live patching where the distribution supports it
- Agree a monthly maintenance window for reboots and test it once
- Send a one-page monthly report listing patches applied and anything still outstanding
- Subscribe to your distributions' security advisories and flag urgent ones to clients
Tools
Risks
Reboots and kernel updates can break old applications, so take snapshots and test first. You take on responsibility for uptime, so put limits of liability in writing. Some clients will balk at a monthly fee until they have an incident, so start with those who already pay for hosting or IT support.
Old-kernel device audit for shops, clinics and offices
Many businesses rely on devices that quietly run old Linux kernels: routers, network storage boxes, point-of-sale terminals, CCTV recorders and Android tablets. These are rarely updated and often exposed on the network. Offer a one-off audit that finds these devices, records their software versions and support status, and gives a prioritised list of what to update, isolate or replace. It is a concrete, inexpensive first security step for organisations that cannot afford a full assessment.
- Best for
- IT technicians, network installers and cybersecurity students building a client base
- First step this week
- Run a scan of your own home or office network this week and write up the findings as a sample audit.
Open the full playbook
Launch steps
- Get written permission from the client before scanning anything
- Scan the network to list devices, vendors and open services
- Check each device's firmware version against the vendor's support and update pages
- Rank devices by exposure and age, and recommend update, isolate or replace
- Offer a follow-up visit to carry out the changes
Tools
Risks
Scanning without written permission can be illegal, so always get a signed scope. Firmware updates can brick devices, so leave risky updates to the vendor or the client's approval. Clients may expect you to fix everything for the audit price, so separate the audit fee from remediation work.
Plain-English security advisory digest for small IT teams
As AI tools surface more bugs, the volume of advisories will grow faster than small teams can read. Publish a weekly digest that translates the most important Linux, router and common-software advisories into plain English: what is affected, how urgent it is and what to do. Start free to build an audience of IT managers, then charge for a premium tier with alerts tailored to each subscriber's software list. The value is filtering and clear action, not breaking news.
- Best for
- Security-minded writers, IT trainers and consultants who want a lead-generation channel
- First step this week
- Write a sample issue covering this week's five most important Linux and network-device advisories and share it with ten IT contacts.
Open the full playbook
Launch steps
- Pick a narrow audience, such as IT managers at schools or small hosting firms
- Follow distribution security trackers, CISA's known-exploited list and vendor advisories
- Summarise each item in three lines: affected, urgency, action
- Publish weekly on a fixed day and collect subscriber software lists for the paid tier
Tools
Risks
Getting severity wrong can mislead readers, so link every item to the original advisory and avoid guessing. Free security news is plentiful, so stay narrow and practical. Growth is slow, so treat the digest as a way to win consulting work rather than a business on its own at first.
Tencent launches Hy4, its first model that reads images and operates a computer
Tencent released the full version of Hy4 on November 4, adding image understanding and computer use for the first time, at about $0.90 per million input tokens, with open weights promised but not yet published.
Tencent launched Hy4, the full version of its flagship Hunyuan model, on November 4. It is the first model in the line that can understand images and operate a computer. Tencent says it can read fine detail in pictures, analyse complex charts and reason about visual maths, and can carry out tasks across both command-line and graphical interfaces, with specialised training on tools such as Blender and CAD software. It also supports tool calling for long, multi-step agent tasks. A preview version, released on August 28, handled text only.
Hy4 is available now through Tencent Cloud's TokenHub and on OpenRouter, and inside Tencent products including CodeBuddy, WorkBuddy and the Yuanbao assistant. Tencent lists the API price at about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens, the same as the preview, and is offering 50% off on TokenHub until December 31. SiliconANGLE reports a 1-million-token context window, and OpenRouter's listing shows 65,536 tokens of maximum output, according to Digital Applied. WorkBuddy and CodeBuddy are free to use from November 4 to 15.
Tencent is pitching Hy4 as one of the best open-source models, in a crowded race with DeepSeek, Alibaba, Moonshot and Zhipu. Its headline comparison is a blind evaluation of 579 finance tasks by 213 professionals, where it says Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. Tencent also says the share of answers containing speculative or unsupported assumptions fell from 23.1% in the preview to 5.6%, and that game-development scores rose 19.1% over the preview. A cheaper Hy4 Flash is described as coming soon.
These claims need care. Every benchmark figure comes from Tencent or its partners, and none has been independently checked yet; sources even disagree on the hallucination-rate figures. The 'open-source' label is so far a promise: Tencent says weights and a system card will follow 'in the coming days', and Digital Applied found the Hugging Face page still locked on November 5. The preview was released under the permissive Apache 2.0 licence with 770 billion total and 49 billion active parameters, but Tencent has not confirmed the size or licence of the final model.
For builders, Hy4 adds another low-cost option that can see screenshots and documents and drive software, which is exactly what agents for back-office work need. At under a dollar per million input tokens, it is cheap enough to test on real workloads alongside whatever you use today. The main questions are practical: whether its quality holds up outside Tencent's tests, whether the weights arrive under a commercial-friendly licence, and whether clients are comfortable sending data to a China-based cloud. Self-hosting the weights, once they appear, would answer that last concern for some.
Why it mattersCheap models that can read screens and documents and operate software lower the cost of building practical automation for small businesses, and add price pressure on Western providers.
👀 What to watch
Whether the weights and licence appear as promised, independent benchmark results in the next few weeks, and the price and launch date of Hy4 Flash.
3 opportunities from this story
Paid model bake-offs for software agencies
New low-cost models such as Hy4 arrive every week, and most agencies do not have time to test them properly. Offer a fixed-price comparison: take 30 to 50 real tasks from a client's product, run them through their current model and two or three cheaper alternatives, and report quality, speed and cost side by side. A clear answer such as 'switching these two features saves this much without visible quality loss' is worth paying for. Keep conclusions tied to the client's own tasks, not public leaderboards.
- Best for
- Freelance AI engineers and developers who already integrate LLM APIs
- First step this week
- Build a small test harness on OpenRouter this week and run Hy4 against your current model on 20 tasks from your own work.
Open the full playbook
Launch steps
- Collect representative prompts and expected outputs from the client, with sensitive data removed
- Run each task through three to four models via one API gateway
- Score outputs blind with a simple rubric and record cost and latency
- Write a two-page report with a recommendation per feature
- Offer to implement routing for the features where savings are clear
Tools
Risks
Results can change quickly as models are updated, so date every report and offer re-runs. Clients may not be allowed to send data to some providers, so check their data rules first. Small test sets can mislead, so state sample sizes and uncertainty plainly.
Chart and statement extraction for small finance teams
Accountants, SACCOs and small lenders spend hours retyping figures from scanned statements, PDF reports and charts. Vision-capable models such as Hy4 can read tables and charts into structured data cheaply. Build a simple service that turns a client's document batch into a clean spreadsheet, with a human checking every total before delivery. Tencent's finance focus makes this a natural use case to test, though you should compare it with other models before committing.
- Best for
- Bookkeepers, data-entry businesses and developers serving accounting firms
- First step this week
- Take five anonymised bank statements or reports, run them through Hy4 and one other vision model, and measure how many figures each gets right.
Open the full playbook
Launch steps
- Pick one document type, such as bank statements from the three most common local banks
- Build an extraction prompt that outputs fixed columns and test it on 50 samples
- Add automatic checks that totals and balances add up
- Have a person review flagged rows before delivery
- Price per page and offer a free first batch
Tools
Risks
Financial documents contain personal data, so comply with data protection law (in Kenya, the Data Protection Act) and get client consent before using any overseas API. Extraction errors can be costly, so keep human review and checksum tests. Some clients will not accept a China-hosted model, so keep an alternative provider ready.
Desktop-task automation pilots for back offices
Many small organisations still run old desktop software with no API: accounting packages, school management systems, hospital records tools. Models that can see a screen and click through an interface make it possible to automate repetitive tasks such as entering invoices or exporting reports. Offer a short, fixed-scope pilot that automates one boring task and measures the hours saved. Start with low-risk, read-only or easily reversible tasks so mistakes are cheap.
- Best for
- Automation freelancers, RPA consultants and developers comfortable with agents
- First step this week
- Record one repetitive task you do in a desktop app and try to automate it with a computer-use model in a test environment.
Open the full playbook
Launch steps
- Interview the client and pick one task done at least weekly that takes over an hour
- Set it up in a sandbox or test copy of the software
- Build the agent with clear limits on what it may click or submit
- Run it alongside a human for two weeks and log errors
- Hand over with a simple dashboard of runs and failures
Tools
Risks
Computer-use agents still make mistakes and can click the wrong thing, so never give them unsupervised access to payments or deletions. Software updates can break automations, so price in maintenance. Vendor claims for Hy4's computer use are unverified, so test several models before promising results.
ChatGPT Images 2.6 makes animated GIFs and adds a faster image model to the API
OpenAI released ChatGPT Images 2.6 on November 4, adding animated GIFs, saved reference images, image markup and SVG export, plus two API models, one of which OpenAI says can return an image in as little as six seconds.
OpenAI released ChatGPT Images 2.6 on November 4. The biggest visible change is animated GIFs: you can turn a written prompt or an uploaded picture into a short looping animation, then change its speed and edit individual frames before sharing it. ChatGPT can also now save reference images of you, your pets or recurring characters, so you can reuse a likeness without uploading it each time. Unite.AI reports the update is rolling out to ChatGPT and Codex users on the web, desktop and mobile apps.
For developers there are two new API models. GPT-Image-2.6 Flare is built for speed, and OpenAI says it can generate an image in as little as six seconds. GPT-Image-2.6 Sunburst is aimed at higher quality and more consistent results across a series of edits. OpenAI's announcement also says image generation is 50% faster, and adds image markup, where you draw on a picture to show what to change, and SVG export, which produces editable vector graphics rather than flat pixels. Maverick Frame, citing OpenAI's documentation, reports GIFs of 2 to 64 frames in the API.
On price, Maverick Frame reports that both new models are billed at $30 per million image output tokens, the same standard rate as GPT Image 2.5, which arrived in September. That does not mean images cost the same: developers on OpenAI's forum report that 2.6 can use several times more output tokens than 2.5 at medium quality for the same prompt, which would raise the cost per image. Unite.AI reports that weekly image generation across ChatGPT and the API now exceeds 4 billion, up from 3 billion at the 2.5 launch.
There are open questions. OpenAI has not published GIF limits per ChatGPT plan, and early user reactions on quality are mixed, with some saying 2.6 looks worse than 2.5 on their prompts. There is naming confusion too: one developer quoted API documentation using 2.5 model IDs for Flare and Sunburst, so check the exact model names in OpenAI's docs before you build. Speed figures such as six seconds come from OpenAI and apply at low quality settings, according to Maverick Frame; real speed will depend on settings and demand.
For small businesses and creators, the important shift is that short animations and vector graphics now come from the same tool many already pay for. A café can turn a menu photo into a looping GIF for WhatsApp Status, and a web designer can request an icon as an SVG file to edit and scale. That makes simple motion and vector work cheaper, which squeezes basic design gigs but creates room for people who can produce consistent, on-brand content at volume and manage the batch work through the API.
Why it mattersAnimated and vector content, which used to need a designer or separate tools, is now a prompt away in a mainstream app, changing what small businesses expect to pay for basic visuals.
👀 What to watch
Plan limits for GIFs in ChatGPT, real per-image costs once developers measure token use, and independent quality comparisons between 2.6 and 2.5.
3 opportunities from this story
Monthly GIF packs for local shops and restaurants
Offer small businesses a monthly pack of short animated GIFs for WhatsApp Status, Instagram and their website: a steaming dish, a rotating product, a 'new arrivals' loop. With ChatGPT Images 2.6 you can produce drafts fast from the owner's own photos, then polish and size them. Many owners want movement in their posts but will not learn video editing. Sell consistency and convenience at a modest monthly price.
- Best for
- Social media freelancers, graphic designers and side-hustlers with a ChatGPT subscription
- First step this week
- Make five GIFs from photos of a local business you know and offer them free in exchange for a testimonial.
Open the full playbook
Launch steps
- Collect product or venue photos from the client each month
- Generate GIF drafts in ChatGPT and adjust speed and frames
- Check every frame for warped text, logos or prices
- Export in sizes for WhatsApp Status, Instagram and the web
- Deliver in a shared folder with a posting calendar
Tools
Risks
AI can distort products, prices or logos, so check every frame and never show something the business does not sell. Clients can learn to do it themselves, so add value through planning and posting. Respect image rights: only animate photos the client owns.
Custom SVG icon and illustration sets for web agencies
SVG export means you can generate icons and simple illustrations as editable vector files that scale cleanly on websites and apps. Offer web agencies and app developers custom, consistent icon sets matched to a client's brand, cleaned up by hand in a vector editor. Agencies often buy generic icon packs that do not match the brand, and custom sets used to be expensive. Charge per set and keep the master files organised so you can extend them later.
- Best for
- Designers comfortable with Figma or Illustrator, and developers with an eye for design
- First step this week
- Generate a 12-icon set in one style through the API, clean it up in Figma, and publish it as a portfolio piece.
Open the full playbook
Launch steps
- Agree a style guide with the client: stroke width, corner radius, colours
- Generate drafts as SVG through the API, one icon per request
- Clean and standardise paths in a vector editor
- Test icons at small sizes and in dark mode
- Deliver as an SVG sprite and a Figma library
Tools
Risks
Generated vectors can be messy and need manual cleanup, so price in your editing time. Copyright protection for AI-generated designs is uncertain in many countries, so tell clients plainly. Icons that look too much like a known brand's can cause problems, so avoid prompting with brand names.
Batch product-image pipeline for online sellers
Online sellers on Jumia, Shopify or Instagram need clean, consistent product photos, often for hundreds of items. Using the faster Flare model through the API, build a pipeline that takes a seller's raw photos and returns consistent backgrounds, sizes and simple lifestyle scenes. Charge per image or per catalogue. The speed matters for volume work, but measure actual token costs per image before setting prices, since 2.6 may use more tokens than 2.5.
- Best for
- Developers and e-commerce service providers
- First step this week
- Process 20 product photos from one seller through the API this week and record the cost and failure rate per image.
Open the full playbook
Launch steps
- Agree a house style with the seller: background, angle, lighting, size
- Build a script that sends each photo with the same prompt and settings
- Track tokens and cost per image and set a price with a safe margin
- Have a person reject images that change the product's shape, colour or label
- Deliver in the marketplace's required sizes and file names
Tools
Risks
Altered product images can mislead buyers and break marketplace rules, so keep the real product unchanged and review every image. API costs per image may be higher than expected, so measure before quoting. Model names and prices may change, so build the pipeline to switch models easily.
Get these as a PDF every 3 days
Free. One email every 3 days. Unsubscribe any time.