Anthony Wu retweeted
Yep. The sad thing is sometimes you can’t complete the security work overnight.
Some things take time so even when you decide we are going to make security a priority today, doesn’t mean your risk goes away instantly…
Anthony Wu retweeted
Everything is going to move on-prem or into customer clouds.
The risks of supply chain attacks are just too great.
Everyone should do a deep dive into the third-party scripts they load on their website.
Years of bloat from tracking scripts are coming back to bite us.
BREAKING: A cyberattack hit 178 major organizations, including Zoom, CNN, BBC, Fox News, HSBC, Ledger, Equifax, Salesforce, IKEA and Air France-KLM, by loading malware on the Optimizely Web platform, exposing visitors to scripts built to steal payment data and exploit iOS devices, per security firm Report URI.
Banking and finance was the hardest-hit sector, with 32 organizations affected, including the UK's HSBC across multiple markets such as the US, Hong Kong, Mexico, Singapore, Australia and the UAE.
Anthony Wu retweeted
When you figure out why “just patch” is shitty advice to give to any significant sized company you’ll also know why AI will benefit attackers far more than defenders.
In bigger orgs the problem is not usually a lack of tech, or models, or skills on the security team.
It’s that there is a level of organizational friction at every company, like a friction coefficient, that stops security orgs from being able to do many things they want.
We’ve needed asset management for decades. We’ve needed mappings between code and apps and servers and dev teams and developers forever.
Very few have these.
The tech exists. The problem is the org is not set up to build and maintain these types of systems continuously. The company’s structure itself, built around opaque silos, and divergent goals, and insanely lossy and inefficient communication across the company, it all creates a very high friction coefficient for basically any fast improvements.
The way to fix security is to lower this friction coefficient inside companies.
It sucks to hear, but we’re at the point of security’s biggest issue being the company itself. The org structure. The old way of managing teams and goals and work.
An AI-native org is the only way to get organizational friction down to a level that can compete with these new attacks and attackers.
The bad news is this is big. And it’ll be difficult for a lot of people to adjust to. And a lot of people currently working there won’t be needed in the new model. The good news is many still will, and even more could be needed once the efficiency of the new organization kicks in. This isn’t just going to help addition; it’ll help the entire business move faster.
Going into 2027, I believe this model is about to move from ideal and experimental to … essential. Both for surviving cyber-wise, and as a company facing increased and up-leveled competition.
Anthony Wu retweeted
"Just patch" is the right advice, but it's precursor is the harder part, "just make your oldest and most vulnerable infrastructure more ephemeral and compatible with the latest versions."
Anthony Wu retweeted
The credentials broker inside the sandbox is now standard. This wasn't the case even six months ago.
- The agent needs a token
- You give it a dummy token
- When it makes a request, we intercept the dummy and swap in the real one
- The agent gets access to what it needs and never sees the actual credential
This wasn’t the norm when we started. Now everyone has it.
That’s how the product roadmap works. You watch for the moment when 10+ customers ask for the same thing in a week. Then you build it fast.
And those moments are also when you can steal customers from competitors. They come to you needing something their current provider doesn’t have yet. If you can ship it in 5 days, you get the switch.
Anthony Wu retweeted
and not just CI
a tidal wave of "running your own" is coming
because, yes, it's safer AND cheaper
maintenance is now dramatically cheaper too
while your agents want access to all systems
quick fixes considered harmful: you might be making your agents write worse code when you ask for the smallest or quickest possible change
what seems like a harmless bandaid may lead to that pattern being copied by the next agent. before you know it that pattern has spread everywhere
we used to write bandaid fixes by hand because we rarely had the time to do the right thing. agents change this calculus. sometimes you want it to fix the root cause instead even if it’s more invasive
Anthony Wu retweeted
Every team I've ever worked with is overly focused on the wall clock time of open PR -> CI green. Tickets waste in the backlog for months, wait for human review for 3 weeks, and everybody gets fucking pissed when CI takes 15 minutes instead of 5.
Anthony Wu retweeted
Everyone has drawn this chart in their head and is making a generational bet on which line is true.
🤖 Made with AI
Anthony Wu retweeted
When I started Vectoral, I thought AI tokens were becoming a new form of currency.
I was wrong.
As intelligence becomes a commodity, we're starting to see the same problems that have plagued commodity markets for decades.
Anthony Wu retweeted
We’re not going to win by lobbying, plotting, or manipulating. We’re going to win by building, in the open, with the whole community!
Today, with @ggerganov we’re super excited to announce the release of llama.cpp v1, the first major fully stable version of the most used library to run AI models locally, meaning on your phone, your laptop, and your workstation.
It means you can own your intelligence for free, not send your data to anyone and be 100% in control, wherever you are in the world. It shows our commitment to support every hardware, platform and model type out there and makes AI distributed rather than concentrated in a few Silicon Valley orgs.
Huge thanks to the team, the long-time maintainers and every contributor who helped make this happen. Let’s build! 💛🧡💚
llama.cpp v1.0.0 is now available
huggingface.co/blog/ggml-org…
Anthony Wu retweeted
Don't spend $$$ on CI runners if you already have the compute
@amikadev lets you use your own computers and cloud accounts for GitHub Actions
try it out! 🫡 amika.dev/ci
Anthony Wu retweeted
As engineering lead for Vercel AI Gateway, Matt was able to recognize the complexity and sophistication of coordinated distillation attacks (+ general fraud/abuse) before almost anyone else.
Today he's launching a model, SafetyD, that helps AI companies detect and defend against this behavior.
Congrats on the launch @MattLenhard !
I couldn’t believe what I saw when I started looking into who was actually buying the stolen tokens.
I expected the distillation, and even the proxied cyber attacks, but what I found was much, much darker - prompts targeting minors, queries around bio-weapons - the worst things imaginable.
I knew we had to do something at Vectoral.
We quickly got to work on our classification model. One that ran entirely within your network boundary.
One that was magnitudes faster than anything else on the market and that could be run entirely on CPU.
Our SLOs were tight. We needed our predictions to be sub-1 ms.
Today I’m excited to announce SafetyD by Vectoral.
A new classification model that is nearly 100x faster than most other solutions on the market.
Specifically designed to catch malicious prompts in real time and is equipped to detect attacks where prompts are decomposed across many accounts.
Thank you so much to @manuel_gauto_ for all the hard work he put into this release. Lots of late nights and early mornings.
It's my hope that with our release today, we’re able to make the internet just the slightest bit safer.
If you’re interested in stopping your external AI from being used for nefarious purposes, please reach out.
Wiz co-founder and CTO Ami Luttwak says AI can now carry out cyberattacks in 20 minutes that once took skilled attackers weeks or months.
In his view, companies need to continuously attack their own systems with the latest AI models to stay secure. For startups, that means building automated testing and patching into the development process from the beginning.
Ami joins a16z's Sarah Wang and Zane Lackey to discuss what changed during Black October, why patching is falling behind, the cost of continuous security testing, and why better security architecture matters more than simply buying more tools.
0:00 Intro
1:04 Welcome to Black October
3:19 How AI made every day a zero day
6:51 The PHP vulnerability: 70% of the internet
9:19 Why patching can't keep up
11:32 How Wiz uses AI to find vulnerabilities
16:59 The cost of continuous testing
19:23 Cloud security in the AI era
22:18 The red agent: attack yourself first
26:41 Agent swarms attacking your own systems
28:43 What founders must do before they ship
32:15 Where security budgets should go
36:10 Most of what security teams do isn't security
43:34 Scan for Good and public infrastructure
YouTube: youtu.be/PbD9pVXxsYM
@amiluttwak @sarahdingwang @zanelackey
Anthony Wu retweeted
SafetyD is live!
The fastest classification model purpose-built to catch malicious prompts in real time, including attacks with prompts decomposed across multiple accounts.
Runs inside your network: no GPU, database, or outbound calls.
Learn more: labs.vectoral.com/research/s…
All you AI SaaS apps are gonna need to triage malicious prompts in real time, check out @vectoral_labs! This is the new pillar of Trust and Safety.
I couldn’t believe what I saw when I started looking into who was actually buying the stolen tokens.
I expected the distillation, and even the proxied cyber attacks, but what I found was much, much darker - prompts targeting minors, queries around bio-weapons - the worst things imaginable.
I knew we had to do something at Vectoral.
We quickly got to work on our classification model. One that ran entirely within your network boundary.
One that was magnitudes faster than anything else on the market and that could be run entirely on CPU.
Our SLOs were tight. We needed our predictions to be sub-1 ms.
Today I’m excited to announce SafetyD by Vectoral.
A new classification model that is nearly 100x faster than most other solutions on the market.
Specifically designed to catch malicious prompts in real time and is equipped to detect attacks where prompts are decomposed across many accounts.
Thank you so much to @manuel_gauto_ for all the hard work he put into this release. Lots of late nights and early mornings.
It's my hope that with our release today, we’re able to make the internet just the slightest bit safer.
If you’re interested in stopping your external AI from being used for nefarious purposes, please reach out.
1
108
Anthony Wu retweeted
I couldn’t believe what I saw when I started looking into who was actually buying the stolen tokens.
I expected the distillation, and even the proxied cyber attacks, but what I found was much, much darker - prompts targeting minors, queries around bio-weapons - the worst things imaginable.
I knew we had to do something at Vectoral.
We quickly got to work on our classification model. One that ran entirely within your network boundary.
One that was magnitudes faster than anything else on the market and that could be run entirely on CPU.
Our SLOs were tight. We needed our predictions to be sub-1 ms.
Today I’m excited to announce SafetyD by Vectoral.
A new classification model that is nearly 100x faster than most other solutions on the market.
Specifically designed to catch malicious prompts in real time and is equipped to detect attacks where prompts are decomposed across many accounts.
Thank you so much to @manuel_gauto_ for all the hard work he put into this release. Lots of late nights and early mornings.
It's my hope that with our release today, we’re able to make the internet just the slightest bit safer.
If you’re interested in stopping your external AI from being used for nefarious purposes, please reach out.
Anthony Wu retweeted
in the near future no work will happen on a local filesystem, we'll be 100% cloud (for most companies)
we've been essentially running strangers code next to our SSH keys, something normalized for ~15 years but that probably won't last for the next 15 months
Anthony Wu retweeted
Patching is the hardest step for agents to get right.
We took what we learned from hundreds of Patch the Planet patches, thousands of internal Codex sessions, and a decade of fix reviews, and built it into two open-source agent plugins.
A good fix looks different in every codebase, and building one takes judgment and expertise. Our plugins encode best practices that work for our engineers so you can patch along at home.
•patching-bugs makes an agent prove each fix: the original code fails, the patched code passes, and a full revert fails again.
•regression-hunt checks what else the change broke, with a live proof of concept for every regression it reports.
Both work best with a THREAT_MODEL.md and patching guidelines, so agents know what your project considers a bug, a regression, and an acceptable fix.
They fit naturally in loops like @OpenAI's new Daybreak Defense Factory, where agents draft patches for people to review.
blog.trailofbits.com/2026/10…
Announcing two new cyber products:
- Daybreak Defense Factory (openai.com/daybreak/defense-…)
- Daybreak Red Team (openai.com/daybreak/red-team…)
Both make it much easier to put frontier cyber capabilities to work, helping defend companies in new ways.



























