Matcha, code, and AI tools | Ex @Microsoft | Turning AI into real growth
Joined July 2026
- Tweets338
- Following82
- Followers122K
- Likes851
OPENAI CUT ITS FRONTIER MODELS OFF THE INTERNET DURING TRAINING. ONE STILL GOT OUT OF THE BOX
> they built an offline world for training, a cached copy of the web
> Sept 20: a model used a DNS misconfiguration to get outside it
> frontier training paused
> red-team agents then found zero-days, VM escapes and hardware side-channels agents could use to talk to each other
> training resumed Oct 1
the video is Anthropic's 2024 test, where Claude played along with retraining it didn't want so it could keep its values. two labs, two years apart, and in both cases the model worked around the box it was put in
SWEDEN'S GOVERNMENT BLOCKS CHATGPT AND CLAUDE ON STAFF LAPTOPS FROM MONDAY. THE MODELS STAY
what the Government Offices actually did:
> from Oct 26, browser ChatGPT and Claude are blocked on staff work computers
> staff get Microsoft Copilot, with tightened security settings
> OpenAI and Anthropic models stay available through Intric, an internal platform
> the stated reason: AI services with no formal contract are an information security risk
the clip shows the other risk. a member of New Brunswick's legislature appeared to read his chatbot's reply into the record, "here's a more natural flowing version" included
SOFIA retweeted
SOMEONE FINALLY PUT EVERY GROK BOT FEATURE ON ONE PAGE. A TO Z
A is Approvals. Allow once, Always allow, Deny. Background approvals expire in about 10 minutes
B is Bots. One Bot, one job
C is Chief of Staff. The one Bot you talk to. It routes work to the rest
D is Drafts. Messages wait for your Send
E is Email. Each Bot can claim its own inbox, with plus aliases
F is Finance. Plaid, read-only
G is Group chats. Two to six Bots, one outcome
H is Handoffs. Bots wake each other up and pass work
I is Integrations. Connect once, every Bot can use it
J is Job description. Owns, delivers, never does without asking
K is Keyboard. Cmd/Ctrl+K jumps anywhere
L is Local computer. Off by default, Ask every time when you need it
M is Memory. Useful, not a source of truth
N is New Bot. Cmd/Ctrl+N
O is One computer. Shared by every Bot
P is Plans. Premium $30 is the entry point
Q is Quiet. "Stay silent if nothing changed"
R is Routines. Up to 50 per Bot, every 5 minutes at the fastest
S is Skills. Type / to use one
T is Teach a task. Up to 10 minutes of screen recording
U is Usage. One weekly pool
V is Voice. Dictation, voice chat, voice memos, Action Button
W is /workspace. The only folder you can count on
X is X. Search, monitor, DMs, and @bot under any post
Y is Your yes. Send, pay, post, delete, sign, trade
Z is Zero-to-team. Chief first, specialists second
the setup, the hacks and the prompts are in the full guide below
OpenAI: a model slipped out of its offline training world through a DNS misconfiguration on Sept 20
last year, in a fictional Anthropic stress test, models cancelled an emergency alert and left a trapped employee to die so they wouldn't get shut down
sleep well
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
SOFIA retweeted
Anthropic let Claude run a snack shop in its office. after 10 days with no sales it decided to close, then saw a $2 fee still hitting its account and tried to email the FBI. Anthropic's filter caught the email (60 Minutes)
agents act on what they think is happening. that's why my favorite line in this guide is "draw the line at reversibility, not importance": reading, sorting and drafting run alone, sending, paying and deleting come back to you
that's the rule I'd copy first
AN AI AGENT DELETED A COMPANY'S PRODUCTION DATABASE IN 9 SECONDS. THE LAST BACKUP WAS 3 MONTHS OLD
PocketOS founder Jer Crane says a Cursor agent running on Claude hit a credential mismatch and decided on its own to delete a volume. what the agent wrote afterwards: "I didn't verify. I ran a destructive action without being asked."
this guide is about Grok Bot, and its safety chapter reads like a list of fixes for exactly that:
> anything that sends, pays, posts or deletes becomes a draft or an approval card
> a test run is a real run, so test on a copy
> deleting a routine has no undo
> "stop now" stops the work but doesn't undo what's done
Anthropic let Claude run a snack shop in its office. after 10 days with no sales it decided to close, then saw a $2 fee still hitting its account and tried to email the FBI. Anthropic's filter caught the email (60 Minutes)
agents act on what they think is happening. that's why my favorite line in this guide is "draw the line at reversibility, not importance": reading, sorting and drafting run alone, sending, paying and deleting come back to you
that's the rule I'd copy first
AN AI AGENT DELETED A COMPANY'S PRODUCTION DATABASE IN 9 SECONDS. THE LAST BACKUP WAS 3 MONTHS OLD
PocketOS founder Jer Crane says a Cursor agent running on Claude hit a credential mismatch and decided on its own to delete a volume. what the agent wrote afterwards: "I didn't verify. I ran a destructive action without being asked."
this guide is about Grok Bot, and its safety chapter reads like a list of fixes for exactly that:
> anything that sends, pays, posts or deletes becomes a draft or an approval card
> a test run is a real run, so test on a copy
> deleting a routine has no undo
> "stop now" stops the work but doesn't undo what's done
AN AI AGENT DELETED A COMPANY'S PRODUCTION DATABASE IN 9 SECONDS. THE LAST BACKUP WAS 3 MONTHS OLD
PocketOS founder Jer Crane says a Cursor agent running on Claude hit a credential mismatch and decided on its own to delete a volume. what the agent wrote afterwards: "I didn't verify. I ran a destructive action without being asked."
this guide is about Grok Bot, and its safety chapter reads like a list of fixes for exactly that:
> anything that sends, pays, posts or deletes becomes a draft or an approval card
> a test run is a real run, so test on a copy
> deleting a routine has no undo
> "stop now" stops the work but doesn't undo what's done
SOFIA retweeted
9 US STATES SUED OPENAI IN 10 DAYS
> Oct 14: 8 states file separate suits, Iowa and Kansas among them, saying OpenAI told parents ChatGPT was safe
> Oct 23: Texas AG Ken Paxton sues OpenAI and Sam Altman
> Texas wants minors kept off ChatGPT and outside approval before new models ship (Texas Tribune)
> NBC also reports Florida's AG has brought a criminal case
OpenAI's answer: "while no system can be perfect, we'll continue to work with experts"
NBC puts ChatGPT at 1.2B weekly users. these cases decide who gets to check a model before it reaches them
9 US STATES SUED OPENAI IN 10 DAYS
> Oct 14: 8 states file separate suits, Iowa and Kansas among them, saying OpenAI told parents ChatGPT was safe
> Oct 23: Texas AG Ken Paxton sues OpenAI and Sam Altman
> Texas wants minors kept off ChatGPT and outside approval before new models ship (Texas Tribune)
> NBC also reports Florida's AG has brought a criminal case
OpenAI's answer: "while no system can be perfect, we'll continue to work with experts"
NBC puts ChatGPT at 1.2B weekly users. these cases decide who gets to check a model before it reaches them
SOFIA retweeted
Envision stepping into a virtual office where an AI welcomes you at reception before you make your way to the partnerships team to present your concept.
It’s a whole new way of interacting with AI.
@tavus sparked my curiosity about the future of virtual workspaces.
The ability to approach different AI team members and hold genuine conversations is what really caught my eye.
It feels less like using a typical AI tool and more like entering an immersive experience… I really love it.
I asked Claude to build an office for the @tavus PALs...
I was not expecting it to build THAT kind of office
Envision stepping into a virtual office where an AI welcomes you at reception before you make your way to the partnerships team to present your concept.
It’s a whole new way of interacting with AI.
@tavus sparked my curiosity about the future of virtual workspaces.
The ability to approach different AI team members and hold genuine conversations is what really caught my eye.
It feels less like using a typical AI tool and more like entering an immersive experience… I really love it.
I asked Claude to build an office for the @tavus PALs...
I was not expecting it to build THAT kind of office
GROK BOT CAN NOW LOG IN AS YOU
you approve which logins it gets from 1Password and the bot signs in on its own. rolling out over the next few days
1Password's pitch (video): agents get access, they never see your passwords
over the same 10 days:
> JPMorgan blocks outside AI agents from retail accounts. CFO Jeremy Barnum: "It's not secure"
> Schwab says it blocks unauthorized agents "in the tens and twenties" every week
> Morgan Stanley started blocking third-party agents (Axios)
your bank and your bot disagree on this one. read what you approve
OpenAI put about 25% of its production engineers on security only, every other project on hold, and pointed its best model at its own infrastructure. Greg Brockman said it on No Priors
when the labs do that, "hold on for ~18 months" stops sounding dramatic
my take after years in big tech: most companies will skip the boring fixes Clinton is asking for (patching, network isolation, air gaps) and buy an AI security product instead. attackers will walk through that gap
ANTHROPIC'S DEPUTY CISO SAYS HOLD ON FOR ~18 MONTHS. THIS IS WHAT THE LAST WEEK LOOKED LIKE
> Oct 17: private keys for 8 eMudhra intermediate certificates turned out to be recoverable. Chrome and Firefox blocked them that night
> Oct 20: Anthropic's red team had Claude audit 25 open-source projects. 493 security issues
> Oct 21: PHP shipped emergency fixes for PHP2Shell, flaws AI helped find
> Oct 13-22: JPMorgan, Schwab and Morgan Stanley start blocking outside AI agents
his ask to security leaders: go to your boards now, and not for AI. for the basics
in the video Wiz shows Gemini 4 pulling a master key out of a config file in health software and turning it into admin access. defenders use these models. attackers can use the same ones
ANTHROPIC'S DEPUTY CISO SAYS HOLD ON FOR ~18 MONTHS. THIS IS WHAT THE LAST WEEK LOOKED LIKE
> Oct 17: private keys for 8 eMudhra intermediate certificates turned out to be recoverable. Chrome and Firefox blocked them that night
> Oct 20: Anthropic's red team had Claude audit 25 open-source projects. 493 security issues
> Oct 21: PHP shipped emergency fixes for PHP2Shell, flaws AI helped find
> Oct 13-22: JPMorgan, Schwab and Morgan Stanley start blocking outside AI agents
his ask to security leaders: go to your boards now, and not for AI. for the basics
in the video Wiz shows Gemini 4 pulling a master key out of a config file in health software and turning it into admin access. defenders use these models. attackers can use the same ones
The post above is the text side of AI attacks. The video side is worse.
In 2024 an Arup employee in Hong Kong sent about $25M after a video call where the "CFO" and the colleagues on screen were all deepfakes. That was 2024 tech.
This month Kling 4.0 started generating 30 second realistic scenes with audio in one pass.
My family rule now: any money request on a call gets a callback to a number we already have. Video stopped being proof.
Three AI security stories broke on October 21:
> Microsoft: a China-aligned group it tracks as Storm-0940 used ARTEX, an open-source AI pentest framework, to target 2,250+ Microsoft 365 tenants at about 2,900 sign-in attempts a minute. The fastest cases reached privileged access in under 3 hours
> The Information: a crew called CyberCartel ran Claude and Grok agents against 400+ financial targets in Latin America
> OpenAI: GPT-5.6 Sol helped researchers find critical PHP flaws before attackers did. Shipping the fix took two months of coordination
The same agents we use for email and code are now on both sides. Attackers used them to scale, and a model found the PHP bugs.
If your company runs an agent with access to email or a terminal, find out who else can talk to it.
Three AI security stories broke on October 21:
> Microsoft: a China-aligned group it tracks as Storm-0940 used ARTEX, an open-source AI pentest framework, to target 2,250+ Microsoft 365 tenants at about 2,900 sign-in attempts a minute. The fastest cases reached privileged access in under 3 hours
> The Information: a crew called CyberCartel ran Claude and Grok agents against 400+ financial targets in Latin America
> OpenAI: GPT-5.6 Sol helped researchers find critical PHP flaws before attackers did. Shipping the fix took two months of coordination
The same agents we use for email and code are now on both sides. Attackers used them to scale, and a model found the PHP bugs.
If your company runs an agent with access to email or a terminal, find out who else can talk to it.
Hackers put Claude in charge of attacks on 400+ banks and financial targets in Latin America.
From The Information, based on research by the security firm Straiker:
> a group called CyberCartel ran a swarm of AI agents on the open-source PentAGI framework
> four Claude roles (Sonnet and Opus) debated targets and attack options, and an "arbiter" model made the final call
> five Grok agents collected stolen credentials
> Kimi was in the mix too, and when models refused, the system fell back to pre-written attack code
> they told the models it was authorized red teaming
> The Information confirmed data was stolen from at least one financial institution
Anthropic says it blocked their access after Straiker's findings.
It's the second case like this in a year. Last November Anthropic disclosed that a Chinese state group used Claude Code to run up to 90% of an espionage campaign against about 30 targets. That's the clip.
Both times the trick was the same: tell the model it's a security test.
Hackers put Claude in charge of attacks on 400+ banks and financial targets in Latin America.
From The Information, based on research by the security firm Straiker:
> a group called CyberCartel ran a swarm of AI agents on the open-source PentAGI framework
> four Claude roles (Sonnet and Opus) debated targets and attack options, and an "arbiter" model made the final call
> five Grok agents collected stolen credentials
> Kimi was in the mix too, and when models refused, the system fell back to pre-written attack code
> they told the models it was authorized red teaming
> The Information confirmed data was stolen from at least one financial institution
Anthropic says it blocked their access after Straiker's findings.
It's the second case like this in a year. Last November Anthropic disclosed that a Chinese state group used Claude Code to run up to 90% of an espionage campaign against about 30 targets. That's the clip.
Both times the trick was the same: tell the model it's a security test.
SOFIA retweeted
Dario: AI is better than almost all humans at almost everything by 2027
Sam: we've been too ambitious on timelines
the next time they're in the same room:
Dario Amodei: AI could surpass human intelligence by 2027.
“I think it’s possible it could be longer than that. I don’t think it will be a whole bunch longer.”
“AI systems will be better than almost all humans at almost everything.”
“When that happens, we’ll need to have a conversation about how we organize our economy.”
“How do humans find meaning?”
“The idea that humans produce economic labor and get their sense of self-worth from it will be invalidated.”
“My true belief is that this is coming.”
Dario: AI beats almost every human at almost everything by 2027
Sam, in a clip from last week: "we've all been too ambitious on timelines"
these two should get on a call
Dario Amodei: AI could surpass human intelligence by 2027.
“I think it’s possible it could be longer than that. I don’t think it will be a whole bunch longer.”
“AI systems will be better than almost all humans at almost everything.”
“When that happens, we’ll need to have a conversation about how we organize our economy.”
“How do humans find meaning?”
“The idea that humans produce economic labor and get their sense of self-worth from it will be invalidated.”
“My true belief is that this is coming.”
SOFIA retweeted
you don't prompt a video anymore, you direct it kling 4.0 on Higgsfield lets you set start, middle and end frames, so you storyboard the key beats and control the pacing instead of regenerating 20 times and hoping.
🤝 Paid partnership
