# Microsoft Bug Hunting AI Outpaces Engineers

An episode of [AI Chat: AI News on Claude, ChatGPT, Gemini, Grok, Perplexity](https://invisiblenetwork.ai/podcast/1669990482), 29 Jul 2026, 19 min.

Audio: https://content.rss.com/episodes/365073/3033447/ai-chat-podcast/2026_07_29_21_18_56_850aa46c-bb70-4fdb-9bdc-c4dfebe712e1.mp3

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! In this episode, we explore how Anthropx Mythos AI is discovering Microsoft bugs at a rapid pace, raising concerns about security vulnerabilities. We also discuss the implications of Google's durable SynthID watermark, the emergence of Encore AI's groundbreaking voice agents, and the changing landscape of writing amidst AI advancements. Chapters 00:00 Microsoft Bug Hunting AI 02:01 Google's SynthID Watermark 04:59 Encore AI's Voice Agents 10:16 Writers Embrace Typos 14:59 Sierra Acquires Oasis Security 17:01 Recent Industry Headlines Show Links • Get the top 80+ AI Models for $8.99 at AI Box: https://aibox.ai • How I Grow and Scale My Business with AI: https://www.skool.com/aihustle • Get the AI Chat Daily Newsletter: https://www.aichatdaily.com/newsletter

## Transcript

[0:02] Welcome to the AI Chat Podcast. I'm your host,
[0:04] Jaden Schaefer. Every day I cover cutting -edge
[0:07] AI news and talk with the leaders behind it,
[0:09] breaking down what it means for your life and
[0:11] business. Anthropix Mythos AI is finding Microsoft
[0:17] bugs faster than Microsoft's engineers can patch
[0:21] them. Google's SynthID watermark survived 300
[0:24] edits. but fragmentation might actually make
[0:27] this completely useless. We'll get into that.
[0:29] Encore AI just raised $30 million to turn sales
[0:32] call transcripts into voice agents. This is an
[0:35] interesting pipeline as far as the data goes.
[0:38] Writers right now are embracing typos, myself
[0:40] included, and all of the idiosyncrasies basically
[0:43] to fight the AI writing that we're seeing everywhere
[0:46] that has just become, you know, basically a plague
[0:49] on the entire internet. Sayer is going to acquire
[0:52] Oasis security for a billion dollars to police
[0:55] AI agents identities. If you have tasks that
[0:58] you do over and over again using AI tools, I'd
[1:01] love for you to check out the AI box builder
[1:03] platform. This is my own startup at AI box dot
[1:06] AI, and it allows you to link together multiple
[1:08] AI models and put prompts in and automate entire
[1:11] processes of repetitive tasks. So you don't have
[1:14] to do them and, you know, put in your prompts
[1:16] and copy and paste between different documents
[1:18] or different AI models over and over again. Cool
[1:20] thing is that all the different AI models, we
[1:23] have over 80 on the platform, work together.
[1:25] So you can have 11 labs creating audio. You can
[1:27] have Google VO3 creating video. You can use ChatGPT
[1:30] to generate images, and you can use Cloud for
[1:32] your writing. You can build out workflows and
[1:34] really cool tools all on the AI Box Builder platform.
[1:38] It's linked in the description, and it's only
[1:40] $8 .99 a month. In addition, you can chat with
[1:43] all the different AI models in one place for
[1:45] that same price. Let's talk about what's going
[1:47] on with Anthropix Mythos. They have a... Bug
[1:50] Hunting AI, and it found 90 critical and 141
[1:53] different important bugs in Microsoft SharePoint
[1:56] just in April. And this is basically the problem
[1:59] here is that it's finding these bugs faster than
[2:01] Microsoft's engineers can patch them. So Microsoft
[2:04] shipped 600 fixes on July 14th, and their internal
[2:09] meetings show that the company was racing a May
[2:12] 31st deadline before hostile governments build
[2:15] equivalent tools. And I'll also say that this
[2:17] isn't the only thing. Microsoft SharePoint isn't
[2:22] the only thing it's been working on. It has found
[2:24] hundreds of critical bugs in Microsoft. 365,
[2:28] Microsoft Teams, Microsoft Copilot, all of that
[2:31] since the beginning of the year. So I think they
[2:33] have about 300 moderate severity SharePoint flaws
[2:37] that are still queued for patching. So I mean,
[2:40] this is kind of a big deal. And on July of this
[2:43] month, or the 14th of the month, Microsoft released
[2:46] patches for over 600 bugs, only seven were low
[2:50] or moderate severity, and one was already being
[2:52] exploited by hackers in the wild. So I mean,
[2:55] it literally found something that was actively
[2:57] being hacked. The Five Eyes Intelligence Alliance
[3:00] warned in late June last month that the window
[3:03] for defenders to outpace AI -equipped attackers
[3:06] would close within months. Internal Microsoft
[3:09] documents show that it might have already closed,
[3:11] right, because we already have one exploit being
[3:13] used. The real danger, I think, is that Mythos
[3:16] can chain multiple small bugs into working exploits
[3:20] that humans miss. So Vin Nugent, who's a senior
[3:23] anthropic advisor and a former NSA AI chief,
[3:26] said that microsoft's standard triage system
[3:28] which is patch critical flaws first low severity
[3:32] last no longer works when ai can weaponize the
[3:35] bottom of the queue so the race is definitely
[3:38] on and speed alone might not be enough right
[3:41] now we also have google's synth id watermark
[3:44] so basically this is an invisible watermark embedded
[3:46] on ai generated images primarily from google
[3:49] but there's some other a bunch of other companies
[3:51] have kind of all worked together to adopt this
[3:52] not everybody has i think grok if I'm remembering
[3:56] correctly, does not do this, but many others
[3:58] do. So, um, when they were, when they did this,
[4:02] these AI generated images, um, when it had the
[4:05] synth ID watermark, apparently it actually survived
[4:08] 300 rounds of compression and resizing and real
[4:12] testing. And so, and it still was able to detect
[4:16] that an image was AI generated. So people might
[4:18] say like, Oh, look, I'm going to, you know, take
[4:20] this AI generated images image. I'm going to
[4:22] compress it. I'm going to resize it. I'm going
[4:24] to, um, zoom in and like, do you, you can. try
[4:27] to edit it in all sorts of ways. But that watermark
[4:29] is still detectable, which is really cool. And
[4:32] that Google says proves that their tech is way
[4:35] more durable than they were expecting. I think
[4:37] there's still one big problem. And that's Google's
[4:40] detector can't read watermarks for open AI runway
[4:43] or NVIDIA. So I believe if I'm correct, they're
[4:46] working with meta on this, and maybe a couple
[4:48] other players. But even though they're using
[4:50] the exact same technology, OpenAI Runway and
[4:53] NVIDIA, like the detector doesn't go cross platform,
[4:57] which is kind of weird. So SynthID is obviously
[5:00] really impressive. And with those 300 compressed
[5:03] cycles on both fully AI generated and AI edited
[5:07] images. It was only breaking after a 20 % border
[5:11] crop was added on top. That was the only thing
[5:13] that was killing it. Google limits synth ID verification
[5:16] to about 10 image checks per day per user, and
[5:19] they throttle faster lookups on similar images
[5:22] to block attackers from reverse engineering their
[5:25] system. It's kind of like when you, when Anthropic
[5:27] is complaining about everyone going and doing
[5:30] distillation attacks, right? They're like, everyone's
[5:32] asking Anthropic questions, getting the answers
[5:34] and training their AM models on it. Google's
[5:36] worried that people are going to use the verification
[5:38] like hey is this was this generated with synth
[5:40] id to reverse engineer how they're doing i don't
[5:44] know if that's really a long -term play because
[5:45] eventually right you're going to get enough data
[5:47] and figure it out but um for now there it's kind
[5:51] of a closely guarded secret it feels kind of
[5:52] like it's going to be like these captchas how
[5:55] captcha has had to evolve a lot right it used
[5:57] to just be like really simple letters and then
[5:59] it got more complicated and then it got turned
[6:00] into like identify all the bridges in this photo
[6:03] and you know zoom this uh you know spin the three
[6:07] model till it lines up in the image. Like there's
[6:08] all these crazy things that have been invented
[6:10] for captures. This is kind of what I feel like
[6:12] synth ID is going to become opening. I runway
[6:14] and NVIDIA have each built their own versions
[6:17] of watermarks. Um, but they all have a little
[6:19] bit different of implementation. That's what
[6:21] I was saying. Like it doesn't actually go cross
[6:22] platform. And so there's a bit of a fragmented
[6:25] landscape right now. And these detectors, they
[6:27] can only really detect their own watermarks,
[6:30] not everyone else's watermarks. The thing that
[6:32] I'm the most curious about, and I don't actually
[6:34] have a direct answer to this, that, that I I
[6:36] would be curious, though, is if it's actually
[6:38] able to, you know, you go to like Google generate
[6:40] an image, maybe it looks like a great image,
[6:42] just go to an open source model, give it that
[6:44] image and say regenerate this. If it would kind
[6:48] of kill the synth ID on top, and I would guess
[6:51] that it might actually kill it. So For that reason,
[6:54] I mean, half the time I'm generating images,
[6:56] I'm like screenshotting, you know, another image
[6:59] or another graphic or something that I like out
[7:00] in the wild. And I'm like, hey, use this as like
[7:02] an inspiration and change it slightly. And so
[7:05] that's how I use like chat GPT to generate images
[7:07] a lot of times. And I'm worried it's going to
[7:09] be the same thing, which basically makes these
[7:12] image generating synth ID things useless, because
[7:14] for a majority of, you know, bad actors out there,
[7:18] they'll be smart enough to figure that out. And
[7:20] 99 % of people that just use these AI image.
[7:22] generators in a you know responsible or good
[7:25] way well it doesn't really matter and so at the
[7:27] end of the day yeah i'm not sure if it's gonna
[7:29] make a huge difference if i'm you know being
[7:31] honest sorry for the pessimism okay encore ai
[7:34] has just raised 30 million dollars to build voice
[7:36] agents they're all going to be trained on real
[7:38] sales calls that have actually closed deals which
[7:41] is fascinating right because we have all these
[7:43] data sets i have a friend in particular and he's
[7:44] got a company and he's got a bunch of sales people
[7:47] that help close deals for his cleaning company
[7:49] and what's interesting is he uses go high level
[7:51] and every single call is recorded for the whole
[7:54] conversation. So he has transcripts of every
[7:56] single call. And yeah, he knows which of those
[7:58] calls actually ended up closing. You can also
[8:00] even if he wasn't tracking that bit of information
[8:03] in particular, you could usually just go read
[8:05] the transcript or have AI read the transcript
[8:06] and decide which ones were closed. But that is
[8:09] fascinating, right? It's like determining who
[8:11] the top sales reps are, which ones are actually
[8:13] closing, using those transcripts and training
[8:16] models to be more like those. So fine tuning
[8:18] models to actually be good at closing deals.
[8:21] This startup in particular, Encore AI, they're
[8:24] getting their call recordings and they're mining
[8:26] all of that to identify what the top performers
[8:28] say that move customers forward. And so, yeah,
[8:32] this is... in my mind, really, really smart.
[8:35] The annual recurring revenue has grown more than
[8:37] 5x since they started and since they did their
[8:40] seed round, which was about 18 months ago. They
[8:42] now serve over 40 enterprise customers. Most
[8:45] of those are financial institutions that they're
[8:46] servicing right now. This was started in 2022
[8:49] as ESAIT .io, and then they rebranded to Encore.
[8:54] Later on, I think one of their advisors recommended
[8:56] it because ESAIT is not, I don't know, very recognizable.
[9:00] In any case, they... analyze different customer
[9:03] interactions stage by stage to find revenue leaks
[9:06] and friction points. And they feed those insights
[9:08] into both AI agents and human employees for coaching.
[9:12] So what's interesting to me is it doesn't just
[9:14] grab that just like the pure transcript. and
[9:17] try to copy the whole thing. But it breaks it
[9:19] into stages. It's like, okay, we're in, you know,
[9:21] maybe it's like a, I don't know if they do cold
[9:23] calling, but let's say it's like cold calling.
[9:24] It's like, you got to get the person to stay
[9:26] on the phone. And it's like, then you got to
[9:27] get them to be interested in the product. Then
[9:28] you got to give them like a hook. Then you got
[9:30] to like, you know, get rid of any sort of issues
[9:34] they might have with your product. Then you got
[9:35] to sell them on the price, right? Like these
[9:36] are these different stages in the sale that you
[9:39] got to go through. And so it's breaking the transcripts
[9:41] down into all of those different pieces to, you
[9:44] know, identify it that way and kind of break
[9:45] it into chunks. I think this is so fascinating.
[9:48] Encore's edge over Salesforce and a bunch of
[9:50] other CRM giants is that the call recordings
[9:53] live outside of their platform. So financial
[9:55] services is, I think, one of the smartest areas
[9:59] that they could get into for this. Banks right
[10:01] now and also insurers already record every customer
[10:03] conversation for compliance. We've all heard
[10:05] it when we're on the phone and it's like, this
[10:06] call may be monitored for quality assurance purposes,
[10:09] right? And so Encore right now is going to be
[10:11] able to plug into that infrastructure without
[10:12] forcing a renegotiation of how CRM data gets
[10:17] handled across every customer deployment. And
[10:19] if they're able to actually train it on, like,
[10:21] let's say they go to Chase Bank and they're like,
[10:22] hey, Chase Bank, you already have all the transcripts.
[10:25] We're going to use your transcripts. So we understand
[10:28] how your top salespeople actually convert people.
[10:30] Right. So and if they can just keep that data
[10:33] with just Chase Bank and they're not like training
[10:35] off of that and giving it to, you know, Bank
[10:38] of America or something. else, that's a really
[10:40] valuable because Chase Bank gets, you know, the
[10:42] best their salespeople are able to do. And Bank
[10:44] of America gets the best their salespeople are
[10:46] able to do. And they're not worried about that
[10:47] data or that, you know, confidential information
[10:50] leaking. I think this is incredible. And so when
[10:52] you have something like Salesforce, currently
[10:54] Salesforce and others like that are a little
[10:57] bit more centralized, and they don't have those
[10:59] kind of out of the box, you know, custom client
[11:02] solution. So very cool. Right now, writers and
[11:05] editors are deliberately breaking conventional
[11:08] style rules. We're all getting rid of em dashes.
[11:11] We're all adding typos. Well, I mean, I actually
[11:14] just naturally have a lot of typos. So that's
[11:16] just me. My wife always rolls her eyes at me.
[11:18] But this is actually becoming something that's
[11:20] more and more popular. And it's kind of funny
[11:22] because in a weird way with this podcast, I've
[11:24] always kind of had that same mantra. A lot of
[11:27] when I first started this podcast. Prior to this,
[11:30] I actually ran a bunch of 100 % AI generated
[11:33] scripts and speaker podcasts for a different
[11:36] startup I had. And I was just doing it for, you
[11:38] know, lead gen or whatever. And they were kind
[11:40] of like educational episodes, evergreen content,
[11:43] not news like this is. But anyways, from the
[11:45] very beginning, I always knew that this was a
[11:46] possibility. And I always said, look, I got to
[11:49] have a real person, myself, my real voice on
[11:50] this podcast talking because people are going
[11:53] to want to hear my background, my stories, my
[11:55] anecdotes. And even to the point where I was.
[11:59] like pauses and ums and buts and like little
[12:01] i don't know human things that you put in there
[12:04] people want to hear that because they know that
[12:05] this isn't an ai talking to them i for one i
[12:08] mean even if the voice because the voices used
[12:10] to be bad and now they're actually really good
[12:12] like if you use 11 labs v3 the voices are incredible
[12:14] sometimes i can't tell it's a it's a real um
[12:17] or it's just an ai if if i'm not careful and
[12:20] i've had calls i've i've heard music um like
[12:23] i can detect it but guys i'm serious like i'll
[12:26] get I'll listen to a song the other day. I added
[12:28] it to my liked folder. on YouTube Music and realized
[12:32] probably like five listens in. I'm like, this
[12:34] is probably AI generated. So anyways, I think
[12:37] this happens more than we would like to admit.
[12:39] I mean, this stuff is getting a lot better. But
[12:41] my mantra has always been like, look, people
[12:43] are going to want a real human. And this is what's
[12:45] happening with writers now. They're adding literal
[12:47] typos and kind of these messy first person stories
[12:50] to make all of their writing look more unmistakably
[12:54] human. It's not even just like looking human,
[12:56] but unmistakable. Like we want people to know
[12:58] this is human. ChatGPT doesn't output text like
[13:01] this. It doesn't use all lowercase. So this is
[13:03] really interesting. One publication in particular
[13:06] said that they got a thousand pitches in May
[13:08] and almost all of them were AI generated. Pitchfork's
[13:11] editorial head has banned AI from review submissions.
[13:15] That was starting in 2024. And I think that was
[13:17] like two years ago. But novelists are actively
[13:21] avoiding a lot of different things. And I mean,
[13:23] you'll see it on Twitter. It's kind of funny
[13:26] or an X or whatever, where everyone like finds
[13:28] em dashes and old literature or they find like
[13:31] just, you know, funny glitches that Claude always
[13:35] says, you know, it's not this, it's that right?
[13:37] Like all those funny things, and they'll screenshot
[13:39] them from old books. and they're like, oh my
[13:41] gosh, I can't believe Jane Austen was using AI
[13:43] to write. So anyways, there's a lot going on
[13:46] inside of the writing world trying to be more
[13:51] human. Sierra is a $12 billion data security
[13:55] company and they're buying Oasis Security for
[13:57] $1 billion to manage AI agent identities inside
[14:01] of companies. So right now, basically every company
[14:04] is putting out all of these different agents
[14:06] to do work, but they need some tools to track.
[14:09] who those agents are, and what they have access
[14:12] to what they're doing. This is a problem that
[14:14] I think a traditional security software would
[14:17] be tracking humans. But now that we have these
[14:19] agents, they also want to track that. And I think
[14:20] it's interesting, because I've seen some tweets
[14:22] on x where people were like, they're screenshotting,
[14:25] I think it was chat GPT, like 5 .6, or whatever.
[14:29] And It was on cursor. They were using cursor
[14:32] and they said they were working on a project.
[14:35] OK, stop for now. Finish this later. And it made
[14:37] a push after I told it to stop. Why did you do
[14:39] that? And I was like, oh, I just like was doing
[14:41] the push. Anyways, these things happen where
[14:43] these agents, you know, especially if you give
[14:44] them a task, especially if it's more open ended,
[14:46] like, hey, just get X, Y, Z thing done. They
[14:49] do it in their own way. And we don't always know
[14:51] what made the change, what made the edit, who
[14:53] did what. And so you can have some of these issues.
[14:56] Anyways, Oasis was founded in 2022. They raised
[14:58] one hundred and nine. million. They specialize
[15:01] in securing non -human identities and they do
[15:04] it through API, you know, API keys that are logging
[15:08] into corporate systems without any humans there.
[15:10] They just raised $600 million, a $12 billion
[15:13] valuation. And they're now making $150 million
[15:16] a year in revenue, but they're still not profitable.
[15:19] They've spent really aggressively. They have
[15:21] three acquisitions that they just did in quick,
[15:24] like really close together. This is their third
[15:27] acquisition in a few months, which is following
[15:29] them buying Rift and Genie Security. I think
[15:32] it's interesting when you have a company that's
[15:34] able to raise so much money. Obviously, they
[15:35] have a really good team. They have a really good
[15:37] visionary at the top. But when they make all
[15:40] these acquisitions, I mean, A, they're smart,
[15:43] but B, they're growing so fast and they have
[15:46] so much money, they have the bandwidth to do
[15:48] it and not to build it from scratch, which is
[15:50] interesting. There's a huge push right now to
[15:53] own the non -human identity market before a lot
[15:55] of these bigger companies like Palo Alto Networks
[15:57] or Okta are kind of moving into this. And so
[16:00] that's why I think they're jumping on that. Other
[16:02] news that happened this week that I just want
[16:03] to cover the headlines. Google's DeepMind has
[16:06] just shipped Lyria 3 .5 in flow music with some
[16:09] better vocals and lyrics. I've talked about Lyria
[16:12] in the past, not been impressed with it. When
[16:14] it was first came out, it was like only able
[16:16] to generate 30 second clips. Then it became,
[16:18] then it was able to do full songs. And now I
[16:20] think they're getting better at vocals and lyrics.
[16:21] So I think this is going to be a big player,
[16:23] challenging Yudio and Suno and stuff. When it
[16:26] first came out, it wasn't much to talk about,
[16:28] but I think Lyria is getting better. I still
[16:30] like Suno more because they have the entire AI
[16:32] studio and they have like music covers you can
[16:34] do. And there's just so much to do with Suno.
[16:36] But I am excited that the underlying model behind
[16:40] Google is getting better. Panagram just raised
[16:43] $9 million to detect AI generated text. It apparently
[16:45] has 99 % accuracy. It's interesting, you know,
[16:47] in light of what we were talking about with everyone
[16:49] using more, I don't know, typos or whatever you
[16:52] want to call it to seem less AI generated. So
[16:56] now we have these tools that are really good.
[16:58] It's interesting because OpenAI in the very early
[17:00] days. came out, they had their own AI text detector.
[17:05] And they actually shut the project down because
[17:07] they said it was too hard to detect AI generated
[17:09] text. Now, I think we all kind of have AI generated
[17:12] text detection built into our brains now, where
[17:16] we're understanding a lot of the phrases that
[17:18] are used, we're understanding em dashes, but
[17:20] those things that AI models are going to be able
[17:22] to get around very easily. I actually have a
[17:25] whole list of things that AI models frequently
[17:27] say that annoy me. And whenever I get AI to write
[17:31] a podcast script or if I get it to go write an
[17:33] article or I go get it to write a blog post or
[17:37] content or like basically any work I'm doing
[17:39] or even like. even websites i very frequently
[17:42] will say like run through this whole list and
[17:44] remove every single m dash from the website every
[17:46] single time that we say x y or z and i just get
[17:49] it to go like run this kind of audit on my content
[17:52] now eventually hopefully the models will just
[17:54] get away from that but uh it is interesting all
[17:58] this ai generated text detection stuff. I think
[18:01] I like I hope that it's good, but I'm always
[18:03] dubious of it because I feel like it's always
[18:05] a cat and mouse game. Okay, there's funny news.
[18:07] Martha Stewart has co founded hint, which is
[18:10] an AI assistant for homeowners. And she has $10
[18:13] million in funding for this project. So that's
[18:15] kind of interesting. If you haven't left a review
[18:17] on the podcast yet, it would help the show out
[18:19] a ton. I would super, super appreciate it. It
[18:22] helps more incredible people like yourself find
[18:24] the show. And so yeah, it's the number one thing
[18:26] that could help this show reach more a bigger
[18:28] audience. If you wouldn't mind if you haven't
[18:30] already done it, it would mean the world to me.
[18:32] But thanks so much for tuning in. I hope you
[18:34] guys have an incredible day. Make sure to go
[18:36] check out AI box .ai if you want to get 80 different
[18:39] AI models all on one platform, and I'll see you
[18:41] in the next episode.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.