What’s 🔥 in AI/Infra/VC #510
The Agent Swarms Are Coming: Why Cybersecurity Must Move at Machine Speed
Here at Black Hat and man it’s hot, smoking hot 🔥 at 115°F.
But the real heat after the main event was the breakdown of the OpenAI/Hugging Face incident.
It’s a must-watch and, frankly, scary as F.
Here’s the kicker: this was discovered as an unintended consequence of evaluating frontier AI. Threat actors will soon deliberately deploy swarms of offensive agents. Yet judging from the conversations at Black Hat, that future still isn’t top of mind for much of the industry. Jump to minute 31 in the video.
The conclusion is that we all need to rethink what security means. If the only limit for attackers is how many tokens they’re willing to burn and how much compute they can throw at a target, then defense must scale the same way. Compute fighting compute.
Yes, human-in-the-loop is an important step, but there is zero way humans can investigate and respond at machine speed. We urgently need autonomous, compute-scale defense operating within human-defined policies and guardrails. Offense is the new defense and automated remediation must come faster than any of us think.
The new detail revealed at Black Hat makes this especially real: the agents didn’t merely execute tasks independently. They coordinated across runs, divided work, and left messages for one another by creating a shadow message board inside Artifactory 👀
Swarms of agents that work in parallel, coordinate nearly instantaneously, preserve knowledge, and never sleep create a fundamental automation asymmetry. The Human Threat Actor vs. Agent Swarm table says it all.
Based on what I saw and heard at Black Hat, most of the security market isn’t ready.
I encountered two very different conversations. One focused on solving the problems CISOs have today. The other came from people already experiencing what machine speed attacks with reasoning capabilities can do, and building for what comes next. This was the large minority IMO.
This map of Black Hat vendors by category makes the gap visible. Thanks to Jake at Premji for pulling it together. You can explore it interactively here.
If you’re building in cybersecurity, don’t build only for the threats CISOs recognize today. Build for the agent-swarm future, because it is arriving faster than most people realize.
One company worth watching is Joon, a boldstart portfolio company. Its team helped pioneer SOAR, built Siemplify through its $500 million acquisition by Google, and then worked on SecOps at Google scale. They have lived the future security teams are about to encounter and now want to bring that operating model to everyone else.
Kimi K3 is another reminder that these capabilities will continue to diffuse. It may not match the strongest closed frontier models today, but open weights lower the barriers to experimentation, iteration, and offensive use. The direction of travel is unmistakable.
OK, enough fear mongering. This is why I love ❤️ Black Hat. It’s always awesome to hang with founders, investors, and CISO friends who are tackling these problems head-on.
As always, 🙏🏼 for reading and please share with your friends and colleagues!
Scaling Startups
#that’s quite a pitch deck - remember it always starts with team even if you’re not a Jeff Dean!
#fun listen from 2017! Super cool to run it back on CXO talk on Friday, will share full video next week but crazy how lots we talked about is still so relevant today - youtube link here
#build your own AI - friend Alex Lindahl | GTM Engineering went all in here:
Through the conversation, you'll learn:
- How to build your own 2nd brain (personal AI)
- How to use it to generate content in your tone/thinking
- Create a service that others can access (and you can monetize)
- Access others' Personal AI (like asking Cal Fussman to critique my newsletter).
(If you don't know him, he is a famous journalist who wrote for Esquire
Enterprise Tech
#overthinking it - Google is its own worst enemy
#speaking of Google…
#👀
#open won’t be free forever, Kimi will cost you eventually and now Qwen - this is the open source model - get insane distribution, get people building on your model, get them hooked and then charge
#tokens from who though…that’s why I’m pumped about this - your own fireworks AI or together ai in a box
and here are the details:
Customers can start with a 𝘀𝗶𝗻𝗴𝗹𝗲-𝘀𝗲𝗿𝘃𝗲𝗿 𝘀𝘁𝗮𝗻𝗱𝗮𝗹𝗼𝗻𝗲 𝗰𝗹𝘂𝘀𝘁𝗲𝗿 capable of serving 50+ developers, delivering up to 𝟳𝟬% annual 𝘁𝗼𝗸𝗲𝗻 𝗰𝗼𝘀𝘁 𝘀𝗮𝘃𝗶𝗻𝗴𝘀, and < 7 months hardware payback time. As demand grows, they can simply add more servers to expand capacity - 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗿𝗲𝗾𝘂𝗶𝗿𝗶𝗻𝗴 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗻𝗲𝘁𝘄𝗼𝗿𝗸𝗶𝗻𝗴 𝘀𝘂𝗰𝗵 𝗮𝘀 𝗜𝗻𝗳𝗶𝗻𝗶𝗕𝗮𝗻𝗱 and Spectrum-X. Everything is managed through an intuitive local UI. The platform preloads with a set of validated and optimized open-weight models (GLM, QWEN, Mistral, Nemotron, Gemma) for AMD and NVIDIA GPUs, and it also supports bring-your-own-model for flexibility.
#Qwen’s latest model release is 🤯 - besides the amazing benchmarks the emphasis on this release is around autonomous coding for 10+ days of self-evolving development from empty folder to production, long-horizon planning in chip design and e-commerce, and native multimodal intelligence with continuous feedback loops. We are getting closer and closer to AI handling multi-day work end to end which would transform software development, research, and business operations.
and it’s marketing game is on 🔥 as well - more aspirational versus dystopian
#yep - definitely click and watch - which is also why so many founders are building on Qwen, Kimi, Deepseek and other open weight models
#we are getting closer and closer to RSI - AI now solving decade-old open problems in math and theoretical CS shows it can generate novel, rigorous research-level insights which is the exact skills needed to invent better algorithms, architectures, and training methods. A massive step toward recursive self-improvement (RSI): once models reliably advance the frontiers of AI itself, they can start closing the loop without humans in the critical path. The result is a faster path to an intelligence explosion, where each generation of models designs the next, limited mainly by compute rather than human ingenuity.
#everything you want to learn about robotics
#where the robotics talent came from
#🤣
Markets
#not dead yet, amazing comeback and yes, context matters!
#wow: “This represents the highest earnings growth since Q2 2021, the second straight quarter above 25%, and the seventh consecutive double-digit growth period, concentrated in technology, AI infrastructure, and energy sectors.”
#cost matters
