Scout、Edit with Copilotに投げられるのすごいな… Copilot StudioとかでもBrowserUseができればいけるのか…？

🔴 AI agents fail to resist prompt injection attacks in new study Researchers from Nanyang Technological University, ST Engineering, IBM Research, and the University of Illinois Urbana-Champaign tested AI agents powered by GPT-5 and Gemini 2.5-Flash against prompt injection attacks and found none consistently resisted them. • Direct prompt injection attacks succeeded more than 79% of the time • Indirect attacks embedded in web content achieved success rates of 41.67% to 68.16% • 3,168 attack simulations conducted using NanoBrowser and BrowserUse

Windows のCodexがBrowserUseうまく使えてなかったの直ったか……？

昨日の続きで、CodexとClaudeCodeの両方に、BrowserUseを使い、Geminiをコントロールさせてみてる で、BrowserUse、この2つちょっと挙動が違うんだよね。で、Codexのほうは、今日は何度やってもGemini側でエラーが出る。そしてClaudeCodeは一発で成功。 うーん。なんだかな。

If you think the 'It works on my machine' days are behind you, wait until your agents hit you with 'It works in my browser' #browseruse Happy Monday!

just open sourced the browser agent i've been building it controls a real browser, learns from tasks, self-heals when websites change, and runs 24/7 on cron built on browseruse with a full agent layer on top check it out: github.com/sediman-agent/sed… stars appreciated !

I have been using browser harness by @browseruse for the last 2 months and slowly but surely it is getter better at my workflows. I really like the pattern of having domain skills so that next time the harness is used for the same or similar task, it can be done faster and with less tokens. I have heavily customised the harness with custom scripts. The best usecase I found is collab-use. I made this since I like using collab for training and fine tuning models for free on the TPUs. But waiting on the platform is a pain as they don't have any API. collab skills started with a 300 step browser-harness workflow and now I have reduced it down to 70 step for reliably using collab for any general task. Thanks for building this self-healing harness!! @mamagnus00 @sauravpanda

We explored how you can build your own minimal Wolfi-based container images with the OSS Chainguard tools and the "minimal" project Ritvik Arya created. What if the image you need isn't one of the free images and it's not in the "minimal" catalog? That's where a bit of agentic magic can help :-) This PoC agent uses a number of tools including BrowserBase and BrowserUse to reconstruct the Apko config from the official catalog info. Here's an example for the Bun.js image:

we've been doing this for VC clients for a while now, and get 2x-3x data compared with BrowserUse and use 1/2 token compared with CDP. Using Pitchbook, LinkedIn and Twitter to find companies: drive.google.com/file/d/1C2m… Browser Use comparison, we get 3x results with same prompt drive.google.com/file/d/14kI… Using 1/2 token compared to CDP on a Pitchbook prompt (typical for Li, X, PB, CB etc. use cases) drive.google.com/file/d/16Qm… kolay gelsin

codexでサクッとVelop WRT Pro 7のインターネット設定を依頼すると自動でやってくれるようになりました。BrowserUse経由ですが、そもそもssh接続もできるのでOpenWrtベースのコマンドを駆使して結構細かな設定も精度高めにこなしてくれます。claude codeでもできますよ多分。

Bunch of tools including firecrawl, openai image gen, claude for content generation, browseruse for posting, redis based scheduling. I am using LangGraph for connecting them to the main LLM.

A bug shows up in your product Lucent spots it Drops an alert in Slack Someone tags a coding agent in the thread It ships a PR Fixed before anyone knew it was broken This is the workflow @lucent_ai users are actually running today Lucent watches your session replays like a human analyst: friction points, bugs, failed upgrades... Patterns no human has time to sit through Connects to PostHog. Quick set up. Julius, Mastra, BrowserUse & more already on it Every session replay you're collecting and not watching is data you're wasting Lucent can help 🎙️ @RaeAlisa_, Founder, @lucent_ai on @fondocom @thestartpod Big thanks to @NapaMediaLLC for making this episode possible

The stack for vibe-coding success: 1. Vercel 2. Supabase or Neon 3. Mem0 or Supermemory 4. Tasklet 5. Claude code cowork 6. VS Code 7. Codex 8. Loops and Resend 9. Twilio 10. Posthog / Mixpanel 11. Google Workspace 12. Deepwiki 13. Postman 14. Exa / Jina / Parallel / Apifty 15. Browseruse Would you add anything?

you can just DJ things I had to let @OpenAIDevs Codex have aux for the GPT 5.5 party trip...here's the playlist it made with a few manual additions from me And yes, it actually made the playlist with @ browseruse. All I did was prompt it music.youtube.com/playlist?l…

Quick context first: this dropped today alongside Parallel Web Systems' Series B announcement (Sequoia-led, $100M at a $2B valuation, more than double their Series A from five months ago). Parag Agrawal's pitch is that agents will use the web a thousand times more than humans ever have, and most of that work will happen in the background. Parallel sells the search/extraction/monitoring APIs that power those agents. Now the chart itself. The reference is doing a lot of work The 1977 ARPANET logical map is one of the most iconic diagrams in tech history. It showed every node on the early internet, mostly universities and labs (MIT, SRI, Harvard, BBN), connected by lines. By echoing that exact aesthetic (the typewriter font, the labelled boxes, the topological layout, even "FIGURE 1."), Parag is making a claim without saying it out loud: Parallel is to the agent web what ARPANET was to the internet. The substrate. The thing the future runs on. The headline reinforces it: "the frontier builders are bringing it to life, pioneers of the future they think should exist." That's vintage tech romanticism, deliberately. He's casting the customers as pioneers and Parallel as the connective tissue. What the categories tell you Instead of nodes being institutions like in 1977, here they're verticals. Each cluster is a category of agentic work, with the standout companies in that category placed on the line: Legal: Harvey, GC AI, Patent Watch, Finch Legal GTM: Attio, Clay, Actively, Monaco, Nooks Coding: Mozilla, Greptile, Strawberry, Browseruse, Amp Code, Macroscope Finance: Rogo, Taktile Marketing: Profound, Airops Data & analytics: Hex, Omni Infra: Baseten, Modal Productivity: Notion, Gumloop, Rocket Consulting: Bridgetown Research PE: Garnett Station Real estate: Opendoor, Build Insurance: Genpact (interesting flex, a publicly traded BPO) Restaurant: Owner Research: Jenni AI Government: Starbridge, Pursuit Science: Convoke The press release confirms the named customers: Harvey, Attio, Modal, and Rogo. So the chart isn't aspirational, these are real users (well, the named ones at least). The [REDACTED] move Roughly 14 nodes are blacked out. That's deliberate. It does three things at once: signals stealth-mode customers under NDA (especially the two AI LAB redactions, which most readers will assume are frontier labs), creates "what am I missing" curiosity, and protects the chart from looking thin in any vertical. It's the same trick a16z and Sequoia use in market maps. Notice the most prestigious slots, AI LAB and SOCIAL NETWORK, are fully redacted. That's a tell. The two flourishes at the edges Bottom right has "100,000 DEVELOPERS AND AGENTS" with little radiating lines, and "JOIN THE NETWORK" as a CTA. So it's not just a brag. It's a recruitment poster. The chart functions as a fundraise asset, a customer logo wall, a developer marketing piece, and a category-defining manifesto in one image. Why founders should pay attention to the craft, not just the content This is a near-perfect category creation move. Instead of saying "we're a search API company," Parallel is reframing the category to "the parallel web," a place that exists, has citizens, has a map. Once you accept the metaphor, every competitor (Tavily, Exa, Firecrawl) has to either fit on Parallel's map or argue against the entire frame. That's positioning judo. The 1977 reference also lets them claim history before they have it. The map doesn't argue Parallel is foundational, it assumes it, and asks you to disagree. The honest read: it's a great piece of fundraise theatre. Whether the substrate claim holds up is a separate question (Tavily and Exa have plenty to say about that), but as a piece of narrative engineering it's worth studying.

MCP servers out of control? I had this once, after some investigation it turned out I had 1000s of browseruse processes and node processes all lingering around from MCP servers for subagents. I started using github.com/steipete/mcporter after that and never had this happen again

Codex and ChatGPT 5.5 with BrowserUse and ImageGen are great. This is the result I got when I told it to copy the BrowserBase landing page. The ability to use ImageGen inside of Codex wild