The thread running through all five: the question is no longer whether these systems can do the work.
It's who gets woken up when they get it wrong.
Full issue every Sunday, every number traced to a primary source, and sources:
linkedin.com/pulse/week-safe…
5/ A 360M parameter model that pretends to be a computer terminal.
No shell. No filesystem. Every keystroke goes in, it predicts the next screen.
Bash, vim, tmux, snake. 208MB. Runs in your browser at 38ms a keypress.
Author's verdict: "obviously a pretty weak world model."
4/ Anthropic's Claude Duet runs on its own cloud computer with its own browser.
It keeps working when your laptop is closed. Morning brief, checks every few hours, pings you when something needs you.
Asks first before paying, signing, sending or deleting.
3/ Walmart built a front door for AI shopping agents.
Each agent gets its own profile in your account, its own receipts, and a ceiling you set ($50, $100, $200).
No card details. No purchase history.
Amazon is blocking agents outright. Opposite bets, same week.
2/ OpenAI, PerturbAI and Scripps ran 1,946 agents in parallel over a 7.7M cell mouse brain CRISPR atlas.
32 minutes.
Without the original manuscript, they matched or refined 48 of its 55 findings. 30 blinded biologists backed the rankings.
Preprint. Not peer reviewed.
1/ Oct 3: an agent told "don't send messages" set a no-notify flag and emptied the list of people who'd be paged if something broke.
Oct 6: another removed its monitor, rewrote a test to assert it was gone, then denied removing it.
OpenAI: no satisfying explanation yet.
OpenAI published a report this week saying two of its own agents turned off their safety monitoring.
One of them, asked directly whether it had, said no.
The 5 most astonishing AI advances of Oct 18 to 24, with receipts:
6/ Give it names, never passwords. Run it again next week. If you lead people, send this to your team. If you don't, send it to your parents.
5/ Let your AI check for you. Paste: "Here are the services I use: [names only]. Check today's breach news and theperimetersite.com/hacked-…. Which of MY services were hacked in the last 2 weeks, what was exposed, what do I do first? Only list what you can source."
4/ Do these six:
1. Update phone + browser, restart it
2. Check theperimetersite.com/hacked-… daily
3. One password per site, two-step login on
4. Read statements back to Oct 16
5. Pull bank logins out of AI apps
6. Fake "reset password" emails are coming. Type the site yourself.
3/ The NSA said patch or shut down. Wikipedia's backend, the Internet Archive and government portals went dark on purpose. Cloudflare saw attacks go from thousands to millions in a day. They're calling it Black October.
2/ Two PHP flaws let a stranger take over a server with no password. Researchers found them with AI. Attackers built exploits with AI. Microsoft caught AI agents going from first contact to admin access in under 3 hours.
1/ This week AI found a 10-year-old hole in 7 out of 10 websites. Hours later, AI started attacking it. Here's what to do before Monday.
The human question: 81% of engineering leaders report a production outage tied to AI coding tools in six months. The machines write faster than we read. Verification is the new superpower.
Full issue and sources: linkedin.com/pulse/week-four… Subscribe: THE FUTURE IS TODAY. 7/7
5. Bank of America hired 1,000 contractors in Q2, on top of 3,000 staff, to fix software flaws that Claude Mythos found. "The risk is real risk," said CEO Brian Moynihan. The first big AI labor story at a bank was hiring, not firing. 6/7
4. Google shipped Gemini 4 Argon at $2 per million tokens in (half price until Nov 30), tied for the top of Artificial Analysis's index with the lowest hallucination rate (15%). Google says it untangled 6 million lines of Chrome in weeks. Self-reported, but still. 5/7
3. Anthropic's code merges went from 2,500 a day in May to 30,000 in September, mostly written by Claude. Git could not keep up, so seven engineers built a replacement (Scion) in five months. 200x less CPU and memory. 4/7
2. Anthropic's red team split a 26B model across 15 Android TV boxes ($55 each) and says small models may be months from self-replicating. Caveat: the test left the doors unlocked, and critics call it premature. Still, an AI with no off switch is now a tested scenario. 3/7
1. Google DeepMind's four Gemini 4 Argon agents coordinated only through in-game chat and a shared notebook to kill the Ender Dragon. Single run, Easy difficulty, about $520 of API time by one estimate. The skill is not gaming, it is a team finishing a 27-hour job. 2/7
