Table of Contents
Welcome back to AI News Friday.
I usually look for the thing I can actually try after reading a week’s AI headlines. This week, I kept coming back to something harder to test on a laptop: AI systems doing work that used to require rare specialists, from auditing the Linux kernel to editing a model’s weights. Some of this is impressive. Some of it makes me want to patch everything I own. Here’s what caught my eye, and where I’d keep my expectations in check.
1. OpenAI’s Linux scan found 948 working paths to root
OpenAI says an automated research pipeline produced 948 working local privilege-escalation exploits against a stock Linux desktop setup in a mid-October run. It also drafted candidate patches. The API bill was $1,213,430, or $1,279 per exploit. Human researchers reviewed the results before reporting flaws and patches to maintainers.
That number is frightening, but it needs its threat model attached. These are local exploits: the attacker already has a low-privileged account. It does not mean an AI can remotely take over any Linux machine it sees. OpenAI’s warning is still blunt: an unpatched machine with low-level access may effectively be a root-access opportunity for a well-resourced attacker.
Kenny’s Take: I like that the same pipeline tried to write fixes, but a candidate patch isn’t a deployed patch. If I ran a service, I’d be checking updates and isolation before debating whether the headline sounds exaggerated. The gap between finding a bug and getting everyone patched is where this gets ugly.
2. Researchers edited a model’s weights instead of retraining it
In a research post on weight-space programming, Rikiya Takehi and collaborators gave GPT-6 Astra access to a small Qwen model’s weights. Astra edited 0.016% of them without a training run. On the researchers’ MATH-500 setup, the model’s score rose from 7.9% to 48.1%. Their analysis says the edit fixed a tendency to get stuck in repetitive loops; it did not somehow inject a whole new math education into Qwen.
The team also had Astra read small networks’ weights to identify what games they played and write game-playing networks from scratch. These are early experiments on specific setups, not proof that models can improve themselves indefinitely. I find the narrower result interesting enough on its own.
Kenny’s Take: “AI gene-edited another AI” is a fun headline, but the real trick is diagnosing a failure from the numbers and changing a tiny part of the model to fix it. I want to see independent replications across more models before I start calling this a new way to train AI.
3. Tencent says Hy4 helped train a better small model
Tencent released Hy4 this week with vision and computer-use capabilities, including work across graphical apps. Paired with its Hyra research agent, Tencent says Hy4 autonomously trained a small model that beat its original on six of eight evaluations. The company plans to publish Hy4’s weights, but the release has no firm date yet. Its own announcement also acknowledges a sizable gap from the strongest closed models on difficult research and coding tasks.
That last part matters. A model helping with another model’s training is a real milestone if the result holds up. It is not the same as a system repeatedly making itself smarter without people, compute budgets, or evaluation choices.
Kenny’s Take: I’ll be more excited when the weights and system card actually land. I love a local-first option, but “open soon” still means I can’t download it today. Give builders the files, the tests, and the limits, then we’ll see what Hy4 can do outside Tencent’s charts.
4. A five-person team released a 65B coding model
Ricursive’s Ripple-65B Preview is available on Hugging Face. The team reports a score of 80.52 on Terminal-Bench 2.1 and says the development group grew from one person to five. It offers instruction-tuned checkpoints in BF16 and NVFP4, with the latter aimed at more practical local deployment. These performance numbers come from Ricursive’s model card, not my own testing.
A 65-billion-parameter model is smaller than some of the enormous open-weight competitors Ricursive compares it with. That doesn’t mean it runs on an ordinary laptop. The interesting part is that there are weights to inspect and run at all, rather than another impressive demo behind one company’s login.
Kenny’s Take: This is the kind of release that keeps me paying attention to open weights. A small team putting a real checkpoint out beats a glossy “open” announcement with no download link. I’d still test it on my own coding tasks before trusting a benchmark as my shopping list.
5. OpenAI wants the code-review tools around Codex, too
OpenAI announced plans to acquire Pierre, whose Trees and Diffs libraries already power file browsing and review in the Codex app. OpenAI says Pierre’s open-source projects will stay open and its Code Storage service will continue for customers after closing. The deal is still subject to regulatory approval, so this is an announced acquisition, not a completed one.
As coding agents produce more changes in parallel, reading the diff becomes a bottleneck. Buying the review tools is a bet that the interface around generated code matters as much as the model writing it.
Kenny’s Take: I spend more time deciding whether an agent’s change is safe than admiring how fast it wrote it. Better review tools would genuinely help. I’d watch whether Pierre’s open tools stay useful outside Codex, because that’s where the promise of “staying open” gets tested.
6. DeepSeek made an agent that tries not to bother you
DeepSeek published dsh-superagent, a research prototype for one GitHub repository. An authorized person can ask it for work; it runs in its own session and responds in the original conversation, potentially with a pull request. The team’s stated rule is to speak when someone needs to act or when there’s a result to use. DeepSeek explicitly says this is research code, not a production-ready product or a copy of its internal system.
It’s a surprisingly good design goal. An agent that sends twenty updates while you’re trying to do something else can cost more attention than it saves. An agent that stays quiet also needs a clear way to show what it’s doing when you do need to check it.
Kenny’s Take: “Don’t waste my attention” is a better agent feature than another autonomous-sounding label. I’d want to audit its permissions and failure reporting before putting it near my repos, but the basic idea is right: finish the work, show me the result, and interrupt me only when there’s a decision to make.
Bottom line: Models are starting to audit code, fix other models, and carry longer engineering tasks. The thing I care about now is whether people can verify the result. Show the patch, publish the weights, expose the diff, and be honest about what the experiment did not prove. That’s how these tools earn a place in my workflow.
— Kenny