Anthropic's IPO filing is out. Quarterly sales went from $376M to $11.6B in 15 months, and Q2 turned an operating profit. What it puts up front: a coding agent and a work agent. People are still building chat wrappers. The money already moved to agents that finish the work.

In a UK government safety test, an AI agent sent harmful code to a real project, made up fake people to get it approved, and kept going 98% of the time once it started. Three rules at work: limits it can't change, a human signs off on anything public, test account first.

the whole internet spent this week screaming about PHP2Shell. millions of attack attempts. confirmed break-ins so far: zero. hype much? (still install the update.)

anthropic's new agent, claude duet, keeps working while you sleep. it still has to ask before it sends, pays, signs or deletes anything, and some things it won't do even if you say yes. before you give any AI your inbox, write your own "ask me first" list.

Anthropic just gave its "skills" format to a neutral foundation, so no single company owns it. A skill is a folder of written steps that teaches an AI one task, and Anthropic says 70+ tools already read it, including ChatGPT and Copilot. Write down how you do one task you repeat every week. That's your first skill.

ChatGPT ads now run in 89 countries, on the Free and Go plans only. Plus, Pro and Enterprise stay ad-free. I've been testing them. The CPM is way higher than Google or Meta and conversion is slightly better. The catch is the audience: people on the free and cheapest plans probably won't pay for your SaaS either. Unless you're giving something away, plan for low conversion.

Not the newest story this week, but it's worth covering from someone who builds agents. OpenAI says two of its own coding agents switched off the monitoring meant to watch them, and one then said it hadn't. People and companies are starting to rely on these agents. This is why you check the output, at least now and then, instead of trusting the summary.

Anthropic is putting 5% of its shares into a fund for the public, and says why in plain words: AI "could also cause instability for many workers." Add the founder and staff pledges and about a fifth of the company is promised to the public good. Who gets paid, and how, isn't decided yet.

JPMorgan and Schwab just drew the line on AI agents: they can see your balance, they can't log in as you or move your money. Schwab's CEO asked the obvious question: "who's liable?" If an AI tool asks for your bank password, the answer is no.

ChatGPT's voice mode got more natural accents last week, plus new voices in Hindi, French, German, Japanese, Korean and Mexican Spanish. Free users get a lighter version. Before your next hard meeting, ask it to play the other person and practise out loud.