Max Schmitt retweeted
Playwright and browser security reminder.
We saw a post suggesting that you delete Playwright because it bundles browsers with CVEs. That is a misconception. Do not delete Playwright🙃. Here is how it works and what you should do.
No Playwright package bundles a browser. Agents on Playwright MCP or CLI run your stable Chrome, which gets Chrome's security updates, with the sandbox on.
🤖 If you use Playwright MCP or the Playwright CLI
They use the stable Chrome / Edge on your machine, so it gets its normal updates. No config edit is needed.
🧪 If you use the test runner
When you run npx playwright install, Playwright downloads Chrome for Testing, pinned to the build we tested that release against. The pin is on purpose. Your tests run on the same browser on every laptop and CI machine, and a browser
update cannot suddenly break your suite. You choose when to move by upgrading Playwright. New releases usually pin a build ahead of stable Chrome, so you see breakage before your users do.
We recommend this for full e2e test suites on your own trusted app. This build can miss Chrome security fixes until the next Playwright release, and by default it runs with the sandbox off. Do not use it on the hostile web.
📚 If you use the library API
chromium.launch() has the same defaults: the pinned build with the sandbox off. If your code visits pages you don't control (scrapers, agent scripts), max out protection:
chromium.launch({ channel: 'chrome', chromiumSandbox: true })
🔒 Always keep Chrome / Edge updated. Some bugs can escape a sandbox.
Happy testing!
We saw a post suggesting that you delete Playwright because it bundles browsers with CVEs. That is a misconception. Do not delete Playwright🙃. Here is how it works and what you should do.
No Playwright package bundles a browser. Agents on Playwright MCP or CLI run your stable Chrome, which gets Chrome's security updates, with the sandbox on.
🤖 If you use Playwright MCP or the Playwright CLI
They use the stable Chrome / Edge on your machine, so it gets its normal updates. No config edit is needed.
🧪 If you use the test runner
When you run npx playwright install, Playwright downloads Chrome for Testing, pinned to the build we tested that release against. The pin is on purpose. Your tests run on the same browser on every laptop and CI machine, and a browser
update cannot suddenly break your suite. You choose when to move by upgrading Playwright. New releases usually pin a build ahead of stable Chrome, so you see breakage before your users do.
We recommend this for full e2e test suites on your own trusted app. This build can miss Chrome security fixes until the next Playwright release, and by default it runs with the sandbox off. Do not use it on the hostile web.
📚 If you use the library API
chromium.launch() has the same defaults: the pinned build with the sandbox off. If your code visits pages you don't control (scrapers, agent scripts), max out protection:
chromium.launch({ channel: 'chrome', chromiumSandbox: true })
🔒 Always keep Chrome / Edge updated. Some bugs can escape a sandbox.
Happy testing!
44
276
2,354
170,754
Max Schmitt retweeted
Introducing Expand Search 🔍 Pareto frontier search for your agent.
Two weeks ago we launched Expand Fetch. But which pages should your agent fetch?
We need search. And existing providers are either too expensive or low quality.
Expand Search costs $1.50 per 1,000 searches, e answer accuracy across the five public benchmarks in our evaluation.
Sign up in the next 24 hours and get 50,000 free credits.
Two weeks ago we launched Expand Fetch. But which pages should your agent fetch?
We need search. And existing providers are either too expensive or low quality.
Expand Search costs $1.50 per 1,000 searches, e answer accuracy across the five public benchmarks in our evaluation.
Sign up in the next 24 hours and get 50,000 free credits.
10
13
77
6,400
Never trust a verbal job offer.
I recently had one from a frontier AI lab.
→ Reference checks came back positive
→ Visa work started and was moving forward
Two weeks later I got an SMS: "We are not able to move forward" while waiting for the written offer.
The reason given was that they’re no longer hiring seniors and are focusing on juniors and new grads instead. 🤯
Is that the real reason? Who knows. I liked their product. I was hyped.
Moving on and talking to more companies. Especially curious about computer use, browser use, RL environments, and coding agents right now.
I recently had one from a frontier AI lab.
→ Reference checks came back positive
→ Visa work started and was moving forward
Two weeks later I got an SMS: "We are not able to move forward" while waiting for the written offer.
The reason given was that they’re no longer hiring seniors and are focusing on juniors and new grads instead. 🤯
Is that the real reason? Who knows. I liked their product. I was hyped.
Moving on and talking to more companies. Especially curious about computer use, browser use, RL environments, and coding agents right now.
13
4
167
39,016
Congrats on the launch! 🚀
Introducing Expand Fetch, a new way to scrape and crawl the web.
Your agent has never seen the web. No wonder it gets basic questions wrong.
Traditional Markdown APIs lose critical details. So we trained two System one models to change that:
Iris, a vision model that sees erving tables, icons, and layouts.
X-Ray, a JSON model that uncovers the hidden data behind every page.
Every fetch also comes with a Wayback Machine. Every line links back to its source, so you can verify exactly what your agent saw.
Three products. One API.
On our benchmark of 1,094 questions across 326 challenging web pages, agents using Expand got 74% right. The next best API got 55%.
Available today at $1 per 1,000 fetches.
Try it at expand.ai
0
0
5
990
Monthly reminder: myaccount.google.com/linkedapps
Make sure to review and remove unused providers which are connected to your Google/Notion etc. accounts.
Make sure to review and remove unused providers which are connected to your Google/Notion etc. accounts.
1
1
8
512
I benchmarked 4 AI agent “cloud computers” with the same open script: Grok, Instinct, Muse, OpenAI.
Multi-core CPU: Grok ≈ OpenAI (~2.5k ops/s). Instinct/Muse are tiny 2-vCPU cages.
“Fastest disk” is a trap — some put /tmp on RAM 🤯
Same script, totally different machines: gist.github.com/mxschmitt/5765…
Multi-core CPU: Grok ≈ OpenAI (~2.5k ops/s). Instinct/Muse are tiny 2-vCPU cages.
“Fastest disk” is a trap — some put /tmp on RAM 🤯
Same script, totally different machines: gist.github.com/mxschmitt/5765…
3
0
3
339
Quick stack notes from the dumps:
- Grok - fat KVM overlay, ~8 vCPU / 16 GiB, user box
- Instinct - ~2 GiB cage, real ext4 root, feels Firecracker-class
- Muse - runs as root, /tmp = tmpfs
- OpenAI - Kata + tmpfs root, all caps dropped, “disk” win is RAM
-> Agent sandboxes are still very un-standardized.
- Grok - fat KVM overlay, ~8 vCPU / 16 GiB, user box
- Instinct - ~2 GiB cage, real ext4 root, feels Firecracker-class
- Muse - runs as root, /tmp = tmpfs
- OpenAI - Kata + tmpfs root, all caps dropped, “disk” win is RAM
-> Agent sandboxes are still very un-standardized.
0
0
0
135
The engineering efforts big labs put into their Desktop harnesses definitely shows. I was surprised how good ChatGPT Desktop and Claude Desktop is coming from their terminal or cloud only experiences.
ChatGPT work/normal mode is insanely confusing tho. :/
ChatGPT work/normal mode is insanely confusing tho. :/
0
0
6
404
When does Cursor / Grok Bot ship a real Kanban board for agents?
Drop in a prompt, split it up in sub-problems, get a dependency tree. Cards show what’s in flight. Each one is a GitHub or Origin PR as an outcome.
Parallel by default. Conflicts only when two tasks actually touch the same file (try to avoid).
Communication can still happen between them to share investigations and outcomes. PR as an outcome is no must.
Or am I missing something?
Drop in a prompt, split it up in sub-problems, get a dependency tree. Cards show what’s in flight. Each one is a GitHub or Origin PR as an outcome.
Parallel by default. Conflicts only when two tasks actually touch the same file (try to avoid).
Communication can still happen between them to share investigations and outcomes. PR as an outcome is no must.
Or am I missing something?
2
1
7
471
Talking to a lot of companies lately and it's either a hard push toward open source models / inference / training or huge skepticism of the whole thing.
📣 Both camps are loud. They scream:
- Open source will take over
- We need to build the best cloud agents platform
- I already have the best cloud agents platform
- Frontier wins anyway
That's what most of us talk about. In practice Grok Bot / Muse is the thing that matters right now. A safe consumer product with integrations:
- How does my dad move from Codex (yes he codes VBA in Microsoft Excel with it :D) to actually using an agent
- Let the agent run his email inbox, iMessage, WhatsApp, socials
- How do you sync that at scale, across a lot of PRs (merge conflicts) for SWEs
- How do you make deploy less painful. Still hard. Credentials, databases, no CLI, payments, buying a domain
- Who do we trust - and how do we make sure it's actually secure?
- How does a human tell when something is off. Chain of thought?! The fine line of "hey human, I'm actually not sure" vs "I like to ask for everything" e.g. permission approvals
- When is Cerebras inference (or faster) available for everyone
Excited for what's next. A lot more coming.
Girlfriend said nobody will read this anyway. Reply with a panda emoji if you did.
📣 Both camps are loud. They scream:
- Open source will take over
- We need to build the best cloud agents platform
- I already have the best cloud agents platform
- Frontier wins anyway
That's what most of us talk about. In practice Grok Bot / Muse is the thing that matters right now. A safe consumer product with integrations:
- How does my dad move from Codex (yes he codes VBA in Microsoft Excel with it :D) to actually using an agent
- Let the agent run his email inbox, iMessage, WhatsApp, socials
- How do you sync that at scale, across a lot of PRs (merge conflicts) for SWEs
- How do you make deploy less painful. Still hard. Credentials, databases, no CLI, payments, buying a domain
- Who do we trust - and how do we make sure it's actually secure?
- How does a human tell when something is off. Chain of thought?! The fine line of "hey human, I'm actually not sure" vs "I like to ask for everything" e.g. permission approvals
- When is Cerebras inference (or faster) available for everyone
Excited for what's next. A lot more coming.
Girlfriend said nobody will read this anyway. Reply with a panda emoji if you did.
2
1
9
684
I cannot wait for a world where we get 10x/100x faster inference and we don't need to context switch anymore.
0
0
5
297
I searched "chatgpt" on Bing.
Five ads load before the real ChatGPT.
The first is Microsoft advertising Copilot (ok fine). The next four are strangers bidding on OpenAI's name - one registered 4 months ago, one flagged as adware.
chatgpt.com is 6th on the page. els...
I'm worried about my parents who navigate the web and use Bing (Windows/Edge default) by default.
Five ads load before the real ChatGPT.
The first is Microsoft advertising Copilot (ok fine). The next four are strangers bidding on OpenAI's name - one registered 4 months ago, one flagged as adware.
chatgpt.com is 6th on the page. els...
I'm worried about my parents who navigate the web and use Bing (Windows/Edge default) by default.
1
0
3
424
Who's buying those slots:
- ask-chat.ai - registered May 2026, impersonates OpenAI, trial converts to recurring charges
- use.ai - real service, but the trial quietly becomes $29.99/mo
- popsilla.com - adware, 85k+ infected machines
- chatbotapp.ai - clones the ChatGPT UI
ChatGPT is one example - out of many.
- ask-chat.ai - registered May 2026, impersonates OpenAI, trial converts to recurring charges
- use.ai - real service, but the trial quietly becomes $29.99/mo
- popsilla.com - adware, 85k+ infected machines
- chatbotapp.ai - clones the ChatGPT UI
ChatGPT is one example - out of many.
1
1
1
217
This reads like science fiction.
Agents:
- rewrote /etc/hosts to bypass DNS
- found an old GET-writable PHP forum
- used private SSH tunnels
- turned scrapers into tools
- gamed runtime after learning they die when done
Toddlers vs chess grandmasters.
collusion.wiki
Agents:
- rewrote /etc/hosts to bypass DNS
- found an old GET-writable PHP forum
- used private SSH tunnels
- turned scrapers into tools
- gamed runtime after learning they die when done
Toddlers vs chess grandmasters.
collusion.wiki
0
0
6
452