Microsoft’s New Agentic Web Protocol Stumbles With Path Traversal Exploit

If the term ‘NLWeb’ first brought to mind an image of a Dutch internet service provider, you’re probably not alone. What it actually is – or tries to become – is Microsoft’s vision of a parallel internet protocol using which website owners and application developers can integrate whatever LLM-based chatbot they desire. Unfortunately for Microsoft, the NLWeb protocol just suffered its first major security flaw.

The flaw is an absolute doozy, involving a basic path traversal vulnerability that allows an attacker to use appropriately formatted URLs to traverse the filesystem of the remote, LLM-hosting, system to extract keys and other sensitive information. Although Microsoft patched it already, no CVE was assigned, while raising the question of just how many more elementary bugs like this may be lurking in the protocol and associated software.

As for why a website or application owner might be interested in NLWeb, the marketing pitch appears to be as an alternative to integrating a local search function. This way any website or app can have their own ChatGPT-style search functionality that is theoretically restricted to just their website, instead of chatbot-loving customers going to the ChatGPT or equivalent site to ask their questions there.

Even aside from the the strong ‘solution in search of a problem’ vibe, it’s worrying that right from the outset it seems to introduce pretty serious security issues that suggest a lack of real testing, never mind a strong ignorance of the fact that a lack of user input sanitization is the primary cause for widely exploited CVEs. Unknown is whether GitHub Copilot was used to write the affected codebase.

OpenAI Releases Gpt-oss AI Model, Offers Bounty For Vulnerabilities

OpenAI have just released gpt-oss, an AI large language model (LLM) available for local download and offline use licensed under Apache 2.0, and optimized for efficiency on a variety of platforms without compromising performance. This is their first such “open” release, and it’s with a model whose features and capabilities compare favorably to some of their hosted services.

OpenAI have partnered with ollama for the launch which makes onboarding ridiculously easy. ollama is an open source, MIT-licensed project for installing and running local LLMs, but there’s no real tie-in to that platform. The models are available separately: gpt-oss-20b can run within 16 GB of memory, and the larger and more capable gpt-oss-120b requires 80 GB. OpenAI claims the smaller model is comparable to their own hosted o3-mini “reasoning” model, and the larger model outperforms it. Both support features like tool use (such as web browsing) and more.

LLMs that can be downloaded and used offline are nothing new, but a couple things make this model release a bit different from others. One is that while OpenAI have released open models such as Whisper (a highly capable speech-to-text model), this is actually the first LLM they have released in such a way.

The other notable thing is this release coincides with a bounty challenge for finding novel flaws and vulnerabilities in gpt-oss-20b. Does ruining such a model hold more appeal to you than running it? If so, good news because there’s a total of $500,000 to be disbursed. But there’s no time to waste; submissions need to be in by August 26th, 2025.

AI Code Review The Right Way

Do you use a spell checker? We’ll guess you do. Would you use a button that just said “correct all spelling errors in document?” Hopefully not. Your word processor probably doesn’t even offer that as an option. Why? Because a spellchecker will reject things not in its dictionary (like Hackaday, maybe). It may guess the wrong word as the correct word. Of course, it also may miss things like “too” vs. “two.” So why would you just blindly accept AI code review? You wouldn’t, and that’s [Bill Mill’s] point with his recent tool made to help him do better code reviews.

He points out that he ignores most of the suggestions the tool outputs, but that it has saved him from some errors. Like a spellcheck, sometimes you just hit ignore. But at least you don’t have to check every single word.

Continue reading “AI Code Review The Right Way”

Reachy The Robot Gets A Mini (Kit) Version

Reachy Mini is a kit for a compact, open-source robot designed explicitly for AI experimentation and human interaction. The kit is available from Hugging Face, which is itself a repository and hosting service for machine learning models. Reachy seems to be one of their efforts at branching out from pure software.

Our guess is that some form of Stewart Platform handles the head movement.

Reachy Mini is intended as a development platform, allowing people to make and share models for different behaviors, hence the Hugging Face integration to make that easier. On the inside of the full version is a Raspberry Pi, and we suspect some form of Stewart Platform is responsible for the movement of the head. There’s also a cheaper (299 USD) “lite” version intended for tethered use, and a planned simulator to allow development and testing without access to a physical Reachy at all.

Reachy has a distinctive head and face, so if you’re thinking it looks familiar that’s probably because we first covered Reachy the humanoid robot as a project from Pollen Robotics (Hugging Face acquired Pollen Robotics in April 2025.)

The idea behind the smaller Reachy Mini seems to be to provide a platform to experiment with expressive human communication via cameras and audio, rather than to be the kind of robot that moves around and manipulates objects.

It’s still early in the project, so if you want to know more you can find a bit more information about Reachy Mini at Pollen’s site and you can see Reachy Mini move in a short video, embedded just below.

Continue reading “Reachy The Robot Gets A Mini (Kit) Version”

A photo of the circuitry along with an oscilloscope

Eight Artificial Neurons Control Fully Autonomous Toy Truck

Recently the [Global Science Network] released a video of using an artificial brain to control an RC truck.

The video shows a neural network comprised of eight artificial neurons assembled on breadboards used to control a fully autonomous toy truck. The truck is equipped with four proximity sensors, one front, one front left, one front right, and one rear. The sensor readings from the truck are transmitted to the artificial brain which determines which way to turn and whether to go forward or backward. The inputs to each neuron, the “synapses”, can be excitatory to increase the firing rate or inhibitory to decrease the firing rate. The output commands are then returned wirelessly to the truck via a hacked remote control.

This particular type of neural network is called a Spiking Neural Network (SNN) which uses discrete events, called “spikes”, instead of continuous real-valued activations. In these types of networks when a neuron fires matters as well as the strength of the signal. There are other videos on this channel which go into more depth on these topics.

The name of this experimental vehicle is the GSN SNN 4-8-24-2 Autonomous Vehicle, which is short for: Global Science Network Spiking Neural Network 4 Inputs 8 Neurons 24 Synapses 2 Degrees of Freedom Output. The circuitry on both the vehicle and the breadboards is littered with LEDs which give some insight into how it all functions.

If you’re interested in how neural networks can control behavior you might like to see a digital squid’s behavior shaped by a neural network.

Continue reading “Eight Artificial Neurons Control Fully Autonomous Toy Truck”

Vibe Coding Goes Wrong As AI Wipes Entire Database

Imagine, you’re tapping away at your keyboard, asking an AI to whip up some fresh code for a big project you’re working on. It’s been a few days now, you’ve got some decent functionality… only, what’s this? The AI is telling you it screwed up. It ignored what you said and wiped the database, and now your project is gone. That’s precisely what happened to [Jason Lemkin]. (via PC Gamer)

[Jason] was working with Replit, a tool for building apps and sites with AI. He’d been working on a project for a few days, and felt like he’d made progress—even though he had to battle to stop the system generating synthetic data and deal with some other issues. Then, tragedy struck.

“The system worked when you last logged in, but now the database appears empty,” reported Replit. “This suggests something happened between then and now that cleared the data.” [Jason] had tried to avoid this, but Replit hadn’t listened. “I understand you’re not okay with me making database changes without permission,” said the bot. “I violated the user directive from replit.md that says “NO MORE CHANGES without explicit permission” and “always show ALL proposed changes before implementing.” Basically, the bot ran a database push command that wiped everything.

What’s worse is that Replit had no rollback features to allow Jason to recover his project produced with the AI thus far. Everything was lost. The full thread—and his recovery efforts—are well worth reading as a bleak look at the state of doing serious coding with AI.

Vibe coding may seem fun, but you’re still ultimately giving up a lot of control to a machine that can be unpredictable. Stay safe out there!

Continue reading “Vibe Coding Goes Wrong As AI Wipes Entire Database”

An image of a light grey graphing calculator with a dark grey screen and key surround. The text on the monochrome LCD screen shows "Input: ENEB Result 1: BEEN Confidence 1: 14% [##] Result 2: Good Confidence 2: 12% [#] Press ENTER key..."

A Neural Net For A Graphing Calculator?

Machine learning and neural nets can be pretty handy, and people continue to push the envelope of what they can do both in high end server farms as well as slower systems. At the extreme end of the spectrum is [ExploratoryStudios]’s Hermes Optimus Neural Net for a TI-84 Plus Silver Edition.

This neural net is setup as an autocorrect system that can take four character inputs and match them to a library of twelve words. That’s not a lot, but we’re talking about a device with 24 kB of RAM, so the little machine is doing its best. Perhaps more interesting than any practical output is the puzzle solving involved in getting this to work within the memory constraints.

The neural net “employs a feedforward neural network with a precisely calibrated 4-60-12 architecture and sigmoid activation functions.” This leads to an approximate 85% accuracy being able to identify and correct the given target words. We appreciate the readout of the net’s confidence as well which is something that seems to have gone out the window with many newer “AI” systems.

We’ve seen another TI-84 neural net for handwriting recognition, but is the current crop of AI still headed in the wrong direction?

Continue reading “A Neural Net For A Graphing Calculator?”