Hard Forks and Their Relationship to AI Bug Fixing  🔗 **Source:** [Hard Forks and Their Relationship to AI Bug Fixing](https://share.snipd.com/episode/f5140a37-8e57-46f5-a0ef-df730e7d930c) ## Highlights - **AI Outpacing Patch Cycles In Major Software** - Anthropics Mythos found hundreds of critical SharePoint and Microsoft 365 bugs faster than Microsoft could patch them. - Jaeden Schafer notes Microsoft released 600 fixes July 14 after Mythos found actively exploited flaws and many more queued. Transcript: Jaeden Schafer Let's talk about what's going on with Anthropics 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 14. And their internal meetings show that the company was racing a May 31 deadline before hostile governments build equivalent tools. And I'll also say that this isn't the only thing, you know, like Microsoft, 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 queued 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, right, because we already have one exploit being used. The real danger, I think, is that Mythos 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. ([Time 0:01:31](https://share.snipd.com/snip/04b693f8-0bff-4a6c-8657-da43fe8d2498)) - **Small Bugs Can Be Weaponized By AI** - Mythos can chain multiple low-severity bugs into working exploits humans miss, breaking standard triage ordering. - Jaeden relays a senior anthropic advisor warning that speed alone won't beat AI-equipped attackers. Transcript: Jaeden Schafer So I think they have about 300 moderate severity SharePoint flaws that are still queued 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, right, because we already have one exploit being used. The real danger, I think, is that Mythos 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. ([Time 0:02:17](https://share.snipd.com/snip/0ea88582-3eec-41f1-8e67-757709a11cb8)) - **SynthID Survives Intense Image Distortion** - Google's SynthID watermark survived 300 rounds of compression and resizing and remained detectable in tests. - Jaeden highlights it failed only after a 20% border crop and Google throttles checks to prevent reverse engineering. Transcript: Jaeden Schafer 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 some other, a bunch of other companies have kind of all worked together To adopt this. Not everybody has. I think Grok, if I believe, 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 the 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, you know, take this AI generated images, image, I'm going to compress it, I'm going to resize it, I'm going to zoom in and like, do you 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 S ID 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. ([Time 0:03:25](https://share.snipd.com/snip/1198ecfe-b566-4cc0-8950-e7296f92b041)) - **Watermark Fragmentation Weakens Detection** - SynthID is fragmented: each provider's watermark/detector detects only its own outputs, limiting cross-platform verification. - Jaeden worries open-source regeneration or screenshot-and-regenerate workflows could bypass the watermark entirely. Transcript: Jaeden Schafer 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 S ID 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 synth ID 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 you 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 is this was this generated with synth ID to reverse engineer how They're doing i don't know if that's really a long-term play because eventually right you're going to get enough data and figure it out but um for now there 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 you know zoom this you know spin the 3d model till it lines up in the image like there's all these crazy things That have been invented for captchas this is kind of what I feel like synth idea is going to become opening I 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 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. And if it would kind of kill the synth ID on top, and I would guess that it might actually kill it. So for that reason, I mean, half the time I'm generating images, I'm like screenshotting, you know, another image or another graphic or something that I like out in the wild. And I'm like, hey, use this as like an inspiration and change it slightly. ([Time 0:04:21](https://share.snipd.com/snip/a9305fd1-d519-43ef-9ce2-0760ba97f258)) - **Training Voice Agents On Real Winning Calls** - Encore AI trains voice agents on real sales call transcripts that closed deals to emulate top performers. - Jaeden shares a friend's cleaning business example where recorded calls and transcripts reveal which reps actually convert. Transcript: Jaeden Schafer Okay, Encore AI has just raised $30 million to build voice agents, they're all going to be trained on real sales calls that have actually closed deals, is fascinating, right? Because we have all these data sets. I have a friend in particular, and he's got a company and he's got a bunch of salespeople that help close deals for his cleaning company. And what's interesting is he uses go high level and every single call is recorded for the whole conversation. So he has transcripts of every single call. And yeah, he knows which of those calls actually ended up closing. You can also even if if he wasn't tracking that bit of information in particular, you could usually just go read the transcript or have AI read the transcript and decide which ones were Closed. But that is fascinating, right? It's like determining who the top sales reps are, which ones are actually closing, using those transcripts and training models to to be more like those. So fine tuning models to actually be good at closing deals. This startup in particular, Encore AI, they're getting their call recordings and they're mining all of that to identify what the top performers say that move customers forward. And so, yeah, this is, in my mind, really, really smart. The annual recurring revenue has grown more than 5x since they started and since they did their seed round, which was about 18 months ago. They now serve over 40 enterprise customers. Most of those are financial institutions that they're servicing right now. This was started in 2022 as ESAIT.io, and then they rebranded to Encore. Later on, I think one of their advisors recommended it because ESAIT is not very recognizable. In any case, they analyze different customer interactions stage by stage to find revenue leaks and friction points. And then they feed those insights into both AI agents and human employees for coaching. So what's interesting to me is it doesn't just grab the, just like the pure transcript and try to copy the whole thing, but it breaks it into stages. It's like, okay, we're in, know, the, the, maybe it's like a, I don't know if they do cold calling, but let's say it's like cold calling. It's like, you got to get the person to stay on the phone. And it's like, then you got to get them to be interested in the product. Then you got to give them like a hook. Then you got to like, um, you know, get rid of any sort of, uh, issues they might have with your product. Then you got to sell them on the price, right? Like these are these different stages in the sale that you got to go through. And so it's breaking the transcripts down into all of those different pieces to, you know, identify it that way and kind of break it into chunks. ([Time 0:07:17](https://share.snipd.com/snip/4a92f849-00a5-4bcb-8293-29f9e71851b7)) - **Edge From Private Call Data For Enterprises** - Encore's advantage is using call recordings that live outside CRMs, letting each enterprise keep private models tuned to their top closers. - Jaeden says banks and insurers already record calls, so Encore can plug into existing compliance infrastructure. Transcript: Jaeden Schafer Encore's edge over Salesforce and a bunch of other CRM giants is that the call recordings live outside of their platform. So financial services is, I think, one of the smartest areas that they could get into for this. Banks right now and also insurers already record every customer conversation from plot compliance we've all heard it when we're on the phone and it's like this call may be monitored For quality assurance purposes right and so encore right now is going to be able to plug into that infrastructure without forcing a renegotiation of how crm data gets handled across Every customer deployment and if they're able to actually train it on like let, let's say they go to Chase Bank, they're like, hey, Chase Bank, you already have all the transcripts, We're going to use your transcripts. So we understand how your top salespeople actually convert people, right? So and if they can just keep that data with just Chase Bank, and they're not like training off of that and giving it to, you know, Bank of America or something else, that's a really valuable Because Chase Bank gets, you know, the best their salespeople are able to do and Bank of America gets the best their salespeople are able to do and they're not worried about that data Or that, you know, confidential information leaking. I think this is incredible. ([Time 0:09:32](https://share.snipd.com/snip/068f992a-bc63-4b96-9c79-180d70cc2870)) - **Writers Add Typos To Prove They're Human** - Writers are deliberately adding typos, m-dash avoidance, and messy first-person voice to signal human authorship. - Jaeden explains publications and novelists use these idiosyncrasies because mass submissions are increasingly AI generated. Transcript: Jaeden Schafer Right now, writers and editors are deliberately breaking conventional style rules. We're all getting rid of m-dashes. We're all adding typos. Well, I mean, I actually just naturally have a lot of typos, so that's just me. My wife always rolls her eyes at me, but this is actually becoming something that's more and more popular. And it's kind of funny, because in a weird way, with this podcast, I've always kind of had that same mantra. A lot of when I first started this podcast, prior to this, I actually ran a bunch of 100% AI generated scripts and speaker podcasts for a different startup I had. And I was just doing it for, you know, lead gen or whatever. And they were kind of like educational episodes, evergreen content, not news like this is. But anyways, from the very beginning, I always knew that this was a possibility. And I always said, look, I got to have a real person myself, my real voice on this podcast talking, because people are going to want to hear my background, my stories, my anecdotes. And even to the point where I was saying like pauses and ums and buts and like little, I don't know, human things that you put in there, people want to hear that because they know that this Isn't an AI talking to them. I for one, I mean, even if the voice because the voices used to be bad, and now they're actually really good, like if you use 11 labs v3, the voices are incredible. Sometimes I can't tell it's a it's a real or it's just an AI if if I'm not careful. I've had calls i've i've heard music um like i can detect it but guys i'm serious like i'll get i'll listen to a song the other day i added it to my liked folder on uh on youtube music and realized Probably like five listens in i'm like this is probably ai generated so anyways there's there's i think this happens more than we would like to admit i this stuff is getting a lot better. But my mantra has always been like, look, people are going to want a real human, and this is what's happening with writers now. They're adding literal typos and kind of these messy first-person stories to make all of their writing look more unmistakably human. It's not even just like looking human, but unmistakable. It's like we want people to know this is human. ChatGPT doesn't output text like this. It doesn't use all lowercase. So this is really interesting. One publication in particular said that they got 1000 pitches in May and almost all of them were AI generated. ([Time 0:10:48](https://share.snipd.com/snip/93f673f8-2922-4ecd-bbed-1f049803e978)) - **Buying Identity For Autonomous Agents** - Sierra is acquiring Oasis Security to manage non-human agent identities and track API-driven agent access inside companies. - Jaeden notes Oasis specializes in securing API keys and non-human logins, a fast-growing market before incumbents move in. Transcript: Jaeden Schafer Sierra is a $12 billion data security company, and they're buying Oasis Security for $1 billion to manage AI agent identities inside of companies. So right now, basically every company is putting out all of these different agents to do work, but they need some tools to track who those agents are and what they have access to, what They're doing. This is a problem that I think a traditional security software would be tracking humans. But now that we have these agents, they also want to track that. And I think it's interesting because I've seen some tweets on x where people were like um they were screenshotting i think it was chat gpt like uh 5.6 or whatever and it oh no it was on cursor They were using cursor and they said um they were working on a project okay stop for now i'll finish this later and it made a push after i told it to stop like why'd you do that and i was like Oh i just like was doing the push anyways these things happen where these agents you know especially if you give them a task especially if it's more open-ended like hey just get xyz thing Done they do it in their own way and we don't always know what made the change what made the edit who did what and so you can have some of these issues anyways oasis was founded in 2022 they Raised 195 million. They specialize in securing non-human identities. And they do it through API, you know, API keys that are logging into corporate systems without any humans there. They just raised $600 million, a $12 billion valuation. And they're now making $150 million a year in revenue, but they're still not profitable. ([Time 0:13:37](https://share.snipd.com/snip/3f26eda5-2c05-461c-8edc-7e5253143a2d))