Latest / Today’s AI News / Mythos Leaked to a Discord Group, SpaceX Bets $60B on Cursor, OpenAI Launches Team Agents

Transcript

Mythos Leaked to a Discord Group, SpaceX Bets $60B on Cursor, OpenAI Launches Team Agents · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by Nynex
• 0:02
  Productions. Imagine building a billion
• 0:06
  dollar state-of-the-art bank vault to hold the world's most
• 0:09
  dangerous secrets, only to leave the back door propped open with
• 0:12
  a brick. That's exactly what's happening
• 0:14
  at the bleeding edge of AI development today.
• 0:17
  And well, we have a clear mission for this deep dive,
• 0:20
  which is unpacking the latest developments in AI, navigating
• 0:24
  the enterprise shifts, and delivering actionable workflows
• 0:27
  for you. Yeah, it is.
• 0:29
  It's an incredibly intense landscape to navigate right now.
• 0:32
  The friction between these cutting edge software
• 0:34
  capabilities and, you know, basic human error is just wild.
• 0:39
  OK, let's unpack this. Because we want to start with a
• 0:41
  story that perfectly illustrates that friction.
• 0:43
  We need to look at Anthropic. Right, the mythos situation.
• 0:46
  Exactly. So on April 10th they launched
• 0:48
  this highly restricted cybersecurity model called
• 0:51
  Mythos, and this tool was deemed so potentially dangerous that
• 0:54
  Anthropic flat out refused to release it to the public.
• 0:57
  Which I mean, that makes sense for a dual use tool, right?
• 1:00
  Yeah, they only deployed it to select partners under a highly
• 1:02
  guarded program called Project Glasswing, which literally
• 1:05
  sparked emergency meetings at the White House.
• 1:08
  Because in the right hands, Mythos can patch infrastructure
• 1:11
  vulnerabilities at superhuman speeds, but in the wrong hands,
• 1:15
  it's basically a skeleton key for global networks.
• 1:18
  Right, which makes what happened next completely baffling.
• 1:22
  Within days of its launch, Mythos leaked, and it didn't
• 1:27
  leak to a rival nation like China or Russia.
• 1:29
  It leaked to a private Discord group.
• 1:31
  Yeah, a random Discord server. It's crazy.
• 1:33
  So I have to ask, is this really a failure of advanced AI safety?
• 1:37
  Or is this just an embarrassing failure of basic IT
• 1:40
  cybersecurity One O 1. What's fascinating here is that
• 1:43
  it fundamentally exposes the gap between theoretical AI safety
• 1:46
  and, well, practical IT deployment.
• 1:49
  They didn't use some highly sophisticated 0 day exploit.
• 1:52
  Wait, they didn't hack the neural network?
• 1:54
  No, not at all. They essentially just guessed
• 1:56
  the deployment URL I. Mean how do you just guess a web
• 1:58
  address for a classified model? That sounds like guessing a
• 2:01
  password by typing password 123. It's slightly more methodical
• 2:05
  than that, but honestly not by much.
• 2:07
  The group utilized naming conventions that were exposed in
• 2:10
  a recent totally separate breach of a company called Mercor.
• 2:14
  Oh, so they just pattern matched the internal staging?
• 2:17
  Server Exactly. By understanding that structural
• 2:20
  pattern and you know, combining it with a borrowed contractor
• 2:23
  login, they simply found the side door and walked in.
• 2:26
  Wow. I know the group claims they
• 2:28
  aren't doing anything malicious and that they actually have
• 2:31
  access to other unreleased models too, but still, it
• 2:35
  highlights this compounding problem, right?
• 2:38
  As partner access grows, keeping these powerful systems under
• 2:42
  wraps is just getting harder. Yeah, the models are getting
• 2:44
  exponentially more capable, but the pipelines we use to deploy
• 2:47
  them are vulnerable to the oldest tricks in the book.
• 2:50
  So while Anthropic struggles to keep its powerful tech in house,
• 2:54
  Elon Musk is realizing that, well, sometimes you have to look
• 2:57
  outside to get the power you need.
• 2:58
  Right. Which brings us to the massive
• 3:00
  corporate shifts happening over at SpaceX.
• 3:03
  Yes, SpaceX just announced A partnership with the AI coding
• 3:07
  startup Cursor, and they've locked in an option to acquire
• 3:10
  them for $60 billion before the end of the year.
• 3:13
  With 10 billion of that completely guaranteed Which?
• 3:17
  Is just a staggering amount of money.
• 3:18
  I mean, how does a $60 billion valuation make sense for a
• 3:22
  coding startup? Well, to understand the
• 3:23
  economics, you have to look at the severe compute bottleneck
• 3:27
  that Cursor hit. Michael Chuwell, Cursor CEO,
• 3:30
  openly admitted their Composer models had reached a compute
• 3:33
  ceiling. So they just ran out of hardware
• 3:35
  horsepower? Exactly.
• 3:37
  They were actually set to raise $2 billion at a 50 billion
• 3:40
  valuation before this deal even came along.
• 3:43
  But raising cash doesn't instantly buy you operational
• 3:46
  data centers. Right, because buying and
• 3:48
  assembling massive GPU clusters takes months or even years.
• 3:52
  Yeah, you need cooling, power grid agreements, all of that.
• 3:55
  But SpaceX has Colossus. It's arguably one of the most
• 3:58
  formidable compute clusters on the planet right now.
• 4:01
  So Courser gets the compute they desperately need.
• 4:05
  But why is Musk doing this? I mean, doesn't xai already have
• 4:07
  grok? They do, but the reality is that
• 4:10
  XAI has been losing the coding race all year.
• 4:13
  Grok hasn't kept pace with Anthropic's clod code or open
• 4:16
  AI's codex. So Grok just isn't cutting it
• 4:19
  for complex software engineering, yeah.
• 4:22
  Musk actually poached 2 lead engineers from Cursor last
• 4:25
  month, Andrew Millisch and Jason Ginsburg, and he basically
• 4:28
  claimed the startup wasn't built right the first time.
• 4:31
  It's like trying to build your own race car, realizing your
• 4:33
  engine is a complete dud, and just deciding to buy Ferrari
• 4:36
  outright to level up quickly. That is exactly what this is.
• 4:39
  He swaps an in house build for an outside product, completely
• 4:42
  bypassing IPO red tape, and potentially turns cursor into a
• 4:46
  SpaceX fueled coding lab. Speaking of getting real work
• 4:49
  done, Open AI is making a massive push to be the invisible
• 4:53
  Co worker on your team. We are seeing a huge enterprise
• 4:56
  shift right now. Yeah, this is a significant move
• 4:59
  away from the single user chat interface.
• 5:01
  Right. So Open AI just introduced
• 5:04
  codecs, powered workspace agents in ChatGPT.
• 5:07
  And to be clear, these aren't just those custom GPTS we saw
• 5:09
  last year. No, definitely not.
• 5:11
  Those old ones were basically just prompt wrappers, you had to
• 5:14
  manually drive them. But the old solo user GPTS do
• 5:19
  still work, by the way, and a conversion tool is coming.
• 5:21
  Good to know, but these new ones, these workspace agents are
• 5:25
  shared autonomous bots that tackle multi step team workflows
• 5:29
  across ChatGPT and Slack. And the craziest part is the
• 5:33
  offline triggers. They retain memory, they call
• 5:36
  connected apps and they trigger on schedules even when you are
• 5:39
  offline. Wait, really?
• 5:41
  So you can log off for the weekend and the agent keeps
• 5:43
  working. Exactly inside Open AI right now
• 5:46
  they are using these for real use cases.
• 5:48
  Sales reps use them for deep account research and follow-ups.
• 5:52
  And I read that their accounting department is using them for
• 5:54
  journal entries and reconciliations completely
• 5:56
  autonomously, right? And here's why this actually
• 5:59
  matters for you. Over the last two years, teams
• 6:01
  have just accumulated scattered prompts and half built
• 6:04
  workflows. The initial GPT store didn't
• 6:07
  really stick for teams. Yeah.
• 6:08
  It's too manual. Exactly.
• 6:09
  But this agentic upgrade solves a real enterprise problem by
• 6:13
  unifying those workflows passively.
• 6:15
  But you don't need an enterprise budget to build an automated
• 6:19
  workflow right now. I want to break down a step by
• 6:21
  step voice dictation strategy that actually proves your
• 6:25
  writing over time. Oh I love this one.
• 6:27
  It connects those high level enterprise concepts directly to
• 6:31
  a personal workflow. How does it start?
• 6:34
  So Step 1, you install a dictation tool called typeless
• 6:37
  on your machine. You open a session with a top
• 6:40
  tier model like codecs or Claude code and you dictate quote draft
• 6:44
  an outline for a short internal memo about and then insert your
• 6:48
  topic. So you're just doing a raw brain
• 6:50
  dump basically. Exactly.
• 6:51
  But Step 2 is the crucial step. You have to tell the agent
• 6:54
  quote. Save this as the initial draft
• 6:57
  and do not edit it. Now create a separate working
• 6:59
  draft that we can revise. Right, you are forcing it to
• 7:02
  preserve that untouched baseline.
• 7:04
  Yes, so important. Then step three, you read that
• 7:07
  working draft and you use typeless to just talk to the AI.
• 7:10
  You leave comments on inaccuracies, missing items, or
• 7:14
  you know those generic AI sounding phrases like when it.
• 7:16
  Uses the word delve. Oh I hate delve.
• 7:18
  So for Step 4, you give it the exact prompt quote, rewrite the
• 7:22
  draft using the comments I left, write it in my tone, use my
• 7:27
  verbiage, no M dashes, preserve the core points, but cut
• 7:32
  anything that sounds generic. That is a solid loop, but the
• 7:35
  real pro tip here is what you do with that untouched draft from
• 7:38
  Step 2. Yes, tell them about the final
• 7:40
  step. So at the end of the week, you
• 7:42
  have an agent compare all your untouched initial drafts against
• 7:45
  your final published drafts. So it actually sees the changes
• 7:48
  you made. Exactly.
• 7:49
  It extracts your edits to update your future editorial rules.
• 7:52
  It maps your exact cadence over time.
• 7:55
  That is such a genius way to build a personalized ghostwriter
• 7:58
  and well to support these amazing workflows.
• 8:01
  The underlying models and hardware are evolving at
• 8:03
  breakneck speed. That's rapid fire through the
• 8:06
  quick hits and everything else happening in AI today.
• 8:08
  Sounds good. Let's start with Google's
• 8:10
  massive hardware shift. They just unveiled their 8th
• 8:13
  generation TP US which are built specifically for agent
• 8:16
  workloads. And the big deal here is that
• 8:18
  they are physically separating the training ships from the
• 8:21
  inference ships, right? Yeah, because training a model
• 8:23
  is heavy lifting that takes months while inference requires
• 8:26
  lightning fast real time reactions.
• 8:29
  Separating them optimizes the whole network.
• 8:31
  That makes a ton of sense. Google also revealed A
• 8:34
  staggering stat. They said 75% of their in house
• 8:38
  code is now AI generated which is boosting their security in
• 8:42
  OPS. It's a massive shift in how
• 8:44
  software is built. Here's where it gets really
• 8:46
  interesting. If Google is writing 75% of its
• 8:49
  code with AI, what does that mean for the human software
• 8:52
  engineer for years? I mean, they transition from
• 8:55
  typing syntax to directing high level logic, and the models
• 8:59
  helping them do that are getting insanely efficient.
• 9:01
  Take Alibaba's Quinn team, for example.
• 9:04
  Right, they just open sourced Quinn 3.627 B.
• 9:07
  Yeah, and that 27 billion parameter model actually
• 9:10
  surpassed its own 397 billion parameter predecessor in coding
• 9:14
  benchmarks. Well, so they severely
• 9:16
  compressed the intelligence requirements.
• 9:18
  That's huge. We're also seeing big leaps in
• 9:20
  the space. Odyssey introduced Odyssey 2
• 9:23
  Max, which is a world model that's three times larger and
• 9:26
  topping real time physics benchmarks.
• 9:28
  It's in private beta right now. And ChatGPT images 2 point O
• 9:32
  just launched as well, yeah. Plus, Ideogram launched custom
• 9:36
  models. You can fine tune their image
• 9:37
  generator on just 15 to 100 personal assets to get
• 9:41
  completely consistent brand outputs.
• 9:44
  That's a game changer for marketing teams.
• 9:46
  We also saw Gemini launch its Enterprise Agent platform.
• 9:49
  Though leaning entirely on these enterprise apprised tools has
• 9:52
  risks, Anthropic faced major user backlash this week after
• 9:56
  removing clawed code for some pro tier users.
• 9:59
  Yeah, Anthropic claims it was just a a small test on the sign
• 10:03
  up flow, but it really paralyzed some workflows.
• 10:05
  It definitely did. But you know, if we connect this
• 10:07
  to the bigger picture, the true test of all this hardware and
• 10:11
  model power is what everyday people actually build with.
• 10:14
  It absolutely. It's not just about Silicon
• 10:16
  Rally. Exactly.
• 10:17
  Which brings us to today's community AI workflow from
• 10:20
  reader Bridget H In New Zealand, Bridget created a full app for
• 10:24
  her farm business using Claude and She.
• 10:25
  Didn't have a coding background right?
• 10:27
  None at all. She used Claude to build an app
• 10:29
  that records and tracks stock purchases, sales, deaths and
• 10:33
  even specific health treatments like shearing and drenching.
• 10:36
  That is incredibly niche and useful.
• 10:38
  It really is, even handles all her invoices and attachments.
• 10:41
  I just love this because it shows how AI is transforming non
• 10:45
  digital industries like farming just as much as it's changing
• 10:50
  tech hubs. It really leaves us with a
• 10:51
  provocative final thought to ponder Today.
• 10:54
  We saw teenagers on Discord hacking into Anthropic's most
• 10:58
  restricted model, and a farmer in New Zealand building a
• 11:01
  complex agricultural app from scratch.
• 11:04
  It's. Quite the contrast.
• 11:05
  It raises an important question moving forward.
• 11:08
  Does the true power of AI ultimately lie with the trillion
• 11:11
  dollar companies building these systems, or with the clever
• 11:15
  everyday users figuring out how to exploit and apply them?
• 11:18
  That is the big question. And that wraps up our deep dive
• 11:22
  for today. Please remember to subscribe to
• 11:24
  stay updated on all these crazy shifts, and we invite you to
• 11:27
  rate the show 5 stars if you're finding this helpful.
• 11:29
  Thank you so much for hanging out with us, and we'll see you
• 11:31
  back for tomorrow's episode.