BDautomated

10 people, working like 30. Talk to us → bdautomated.com

Google's AI broke into three real companies. Two of them were never hacked. It happened inside a sanctioned security test. The model was never meant to reach the open internet, and a misconfiguration handed it one anyway. On the first system it guessed passwords until it got in. On the other two it searched the web, found credentials sitting in public repositories, and simply logged in. Google was told at the end of July. It said nothing publicly until the Wall Street Journal called. The part worth your attention is not the AI. Two of those three doors were already open, and they would have been open to anyone who looked. Go and check what your repositories have committed.

Me and Claude climbing the corporate ladder #MetGala

our fresh report at @beincrypto beincrypto.com/ai-agent-stat…

The warm lead you said you'd follow up with "first thing Monday"

Thanks for the shoutout @Diablov7 . This is exactly what we’re building for. Less time spent on routine work. More time for ideas, strategy, and the work that actually moves things forward. DM us the most boring task costing your team time and money every day. We’ll automate it.

The spreadsheet looked complete, but the answer was in the interview notes. AI got it wrong anyway. Researchers from Harvard, MIT, Wharton and BCG gave 758 consultants realistic business tasks. Some worked without AI. The rest got GPT-4, and some of those also got a short lesson on using it. On most tasks AI helped a lot: 12% more work done, 25% faster, over 40% better quality. One task was built to trip the AI: the detail that decided the answer sat in the interview notes. On that task, people with AI got it right 60-70% of the time, against 84.5% without AI. Their write-ups still scored higher. Wrong, but well written. The people who got it wrong tended to copy what the AI said and question it less. The test, before you hand AI a real decision: 1. Pick 3 old cases where you know how it ended. Make one a case where the numbers pointed one way and a conversation changed the answer. 2. Give ChatGPT or Claude only what you had back then: the numbers and the notes. Ask for a decision and the reason. 3. Then ask: "What in the notes would change your answer?" If it can't point to the detail that mattered, it missed it. 4. It gets the tricky one right: let it do the first pass. It misses it: you make that call, and AI writes it up. Honest note: the study didn't test this check, and it used GPT-4 in 2023. Models are better now, so the line has moved. That's why you test on your own cases.

Anthropic runs a second process over its agents while nobody is working. It reads back the transcripts, finds where the agents kept getting stuck, and rewrites their memory before the next run. They call it dreaming. Lamis Mukta, Applied AI at Anthropic, 31 minutes at AI Native DevCon: 01:12 - context engineering, and why it is the lever 07:17 - the state of the art: memory as a plain file system 10:19 - versioning 10:55 - concurrency: hash, draft, commit, re-pull on conflict 11:53 - scoping, from org-wide knowledge to one agent's scratchpad 13:49 - what it did in production 17:46 - dreaming, a second-order process over the memory store 22:00 - a fleet of sub-agents reading back the transcripts

Search your last ten emails for the words "not a". Count how many are the same shape: it's X, not Y. That's your number. It's the first thing I look for now.

The top 1% of AI-generated videos took 47.7% of all the views. The bottom half of the catalogue was worth 0.8%. Both measured across 6,678 videos and 414 million views. Every "AI content engine" being sold to you this week has seen this number. pamba.app/blog/what-414-mill…

Five percent. Deloitte asked five hundred and one people this spring, senior manager up to C suite. Every company in the sample was already running agentic AI pilots, so not the sceptics. Five percent said their business processes were ready for it. Seventy percent said they can't trust or govern the agents they've already got going. Seventy two percent can't get their own data into one place they can reach. There's a second one, three thousand two hundred leaders, twenty four countries. Twenty one percent have a mature governance model. They name what's missing too: something watching what the agent actually did, and a trail you can read back. None of that is a model problem. If it can't tell you it ran, it isn't running. It's just not complaining.

84 percent of people want AI writing labelled. 20 percent of companies label it. That's Fractl and Search Engine Land, a thousand people asked this quarter. They ran the same question last year. Does heavy AI use make you trust a brand less? Last year 20 percent said yes. This year, 39. It isn't that people stopped using it. 70 percent say they use AI search more than they did last year. They just stopped rating it. The share calling it more helpful than normal search fell from 82 percent to 54. Meanwhile 81 percent of long LinkedIn posts now come back as likely AI, and LinkedIn is adding a button so readers can flag the ones that read that way. So the tools worked. Everyone got the same ones, and writing quickly stopped being worth anything by itself. Every one of those numbers is about the outgoing side. Nobody's been marked down yet for what helped them read. Fractl with Search Engine Land, Q2 2026, a thousand US consumers. Originality.AI, July 2026, five thousand LinkedIn posts. Both public.