Exploring Latent Space of AI Bug Resolution

**SPEAKER_1** (0:00)
Anthropix Mythos AI is finding Microsoft bugs faster than Microsoft's engineers can patch them. Google Synth ID Watermark survived 300 edits, but fragmentation might actually make this completely useless. We'll get into that. Encore AI just raised $30 million to turn sales call transcripts into voice agents. This is an interesting pipeline as far as the data goes. Writers right now are embracing typos, myself included, and all of the idiosyncrasies, basically to fight the AI writing that we're seeing everywhere, that has just become basically a plague on the entire Internet. Sierra is going to acquire Oasis security for a billion dollars to police AI agents' identities. If you have tasks that you do over and over again using AI tools, I'd love for you to check out the AI Box Builder Platform. This is my own startup at AIBox.AI, and it allows you to link together multiple AI models and put prompts in and automate entire processes of repetitive tasks, so you don't have to do them and put in your prompts and copy and paste between different documents or different AI models over and over again. And the cool thing is that all the different AI models, we have over 80 on the platform work together, so you can have 11 labs creating audio, you can have Google V03 creating video, you can use Chat GPT to generate images, and you can use Claude for your writing. You can build out workflows and really cool tools all on the AI Box Builder Platform. It's linked in the description, and it's only $8.99 a month. In addition, you can chat with all the different AI models in one place for that same price. Let's talk about what's going on with Anthropix Mythos. They have a bug-hunting AI, and it found 90 critical and 141 different important bugs in Microsoft SharePoint just in April. And this is basically the problem here is that it's finding these bugs faster than Microsoft's engineers can patch them. So Microsoft shipped 600 fixes on July 14th, and their internal meetings show that the company was racing a May 31st deadline before hostile governments build equivalent tools. And to also say that this isn't the only thing, you know, like Microsoft SharePoint isn't the only thing it's been working on. It has found hundreds of critical bugs in Microsoft 365, Microsoft Teams, Microsoft Copilot, all of that since the beginning of the year. So I think they have about 300 moderate severity SharePoint flaws that are still cued for patching. So I mean, this is kind of a big deal. And on July of this month, or the 14th of the month, Microsoft released patches for over 600 bugs. Only seven were low or moderate severity, and one was already being exploited by hackers in the wild. So I mean, it literally found something that was actively being hacked. The Five Eyes Intelligence Alliance warned in late June last month that the window for defenders to outpace AI-equipped attackers would close within months. Internal Microsoft documents show that it might have already closed, because we already have one exploit being used. The real danger, I think, is that Mithos can chain multiple small bugs into working exploits that humans miss. So Vin Nugent, who's a senior anthropic advisor and a former NSA AI chief, said that Microsoft's standard triage system, which is patch critical flaws first, low severity last, no longer works when AI can weaponize the bottom of the queue. So the race is definitely on, and speed alone might not be enough. Right now, we also have Google's Synth ID watermark. So basically, this is an invisible watermark embedded on AI-generated images, primarily from Google, but there's a bunch of other companies have kind of all worked together to adopt this. Not everybody has. I think Grok, if I'm remembering correctly, does not do this, but many others do.
So when they did this, these AI-generated images, when it had this SynthID watermark, apparently, it actually survived 300 rounds of compression and resizing and real testing, and it still was able to detect that an image was AI-generated. So people might say, like, oh, look, I'm going to take this AI-generated image, I'm going to compress it, I'm going to resize it, I'm going to zoom in and try to edit it in all sorts of ways, but that watermark is still detectable, which is really cool. And that, Google says, proves that their tech is way more durable than they were expecting. I think there's still one big problem, and that's Google's detector can't read watermarks for OpenAI, Runway, or NVIDIA. So I believe, if I'm correct, they're working with Meta on this, and maybe a couple other players. But even though they're using the exact same technology, OpenAI, Runway, and NVIDIA, like, the detector doesn't go cross-platform, which is kind of weird. So SynthID is obviously really impressive, and with those 300 compressed cycles on both fully AI-generated and AI-edited images, it was only breaking after a 20% border crop was added on top. That was the only thing that was killing it. Google limits SynthID verification to about 10 image checks per day per user, and they throttle faster lookups on similar images to block attackers from reverse engineering their system. It's kind of like when Anthropic is complaining about everyone going and doing distillation attacks, right? They're like, everyone's asking Anthropic questions, getting the answers, and training their AM models on it. Google's worried that people are going to use the verification like, hey, was this generated with SynthID to reverse engineer how they're doing? I don't know if that's really a long-term play, because eventually, you're going to get enough data and figure it out. But for now, it's kind of a closely guarded secret. It feels kind of like it's going to be like these CAPTCHAs, how CAPTCHA has had to evolve a lot, right? It used to just be like really simple letters, and then it got more complicated, and then it got turned into like, identify all the bridges in this photo, and zoom this, spin the 3D model till it lines up in the image. There's all these crazy things that have been invented for CAPTCHAs. This is kind of what I feel like SynthID is going to become. OpenAI, Runway, and NVIDIA have each built their own versions of watermarks, but they all have a little bit different of implementation. That's what I was saying, like it doesn't actually go cross-platform. And so there's a bit of a fragmented landscape right now, and these detectors, they can only really detect their own watermarks, not everyone else's watermarks. The thing that I'm the most curious about, and I don't actually have a direct answer to this that I would be curious, though, is if it's actually able to, you know, you could go to like Google, generate an image, maybe it looks like a great image, just go to an open source model, give it that image and say, regenerate this.