The automation box becomes the target
An attacker pointed an AI agent at 460 servers and got in through the same self-hosted tools small firms run. Meanwhile the EU Act starts being enforced and the cheap end of the model market falls out of the floor.
- 01
An agent-run attack campaign hit the tools small firms self-host
Unit 42 published a breakdown of a campaign where one operator drove an open-source agent framework over Telegram and let a model do the target-finding and exploit selection. More than 460 internet-facing systems were probed; three organisations had data pulled via a Citrix NetScaler flaw and command execution was confirmed on eleven Marimo notebook instances. The exploited list includes n8n and Langflow — the exact automation tools a small firm sticks on a spare box and forgets about.
Unit 42, Palo Alto Networks - 02
The EU AI Act moves from countdown to enforcement
From 2 August the Commission's AI Office and national authorities started enforcing the Act, with the transparency rules live: an interactive AI has to tell people it is not a human, and generated or altered content has to be labelled and carry machine-readable marks. The Commission also points at its complaints tool and a whistleblower channel — meaning enforcement no longer depends on a regulator noticing you. Over 180 organisations have signed the transparency Code of Practice voluntarily.
European Commission - 03
OpenAI's cheap tier drops to 20 cents per million tokens
OpenAI reduced prices on the two lower GPT-5.6 tiers at the end of July. The pricing docs now list Luna at $0.20 in and $1.20 out per million tokens, Terra at $2 and $12, and the flagship Sol unchanged at $5 and $30 — halved again on the batch and flex routes. If you priced an automation six months ago and shelved it because the sums didn't work, the sums have changed.
OpenAI API pricing - 04
Three open letters, three different bets on how fast this should go
Willison collected the open letters that landed in late July: a Microsoft-organised defence of open-weight models with 235+ signatories including Nvidia, Amazon and OpenAI; Anthropic's separate position published three days later; and 'Pacing the Frontier', signed by 1,324 people who work inside the frontier labs asking the US government to back international efforts to slow automated AI development. Worth ten minutes because it tells you which way the tools you depend on may be pushed.
Simon Willison - 05
Gemini now draws diagrams inside Google Docs
Rolling out from 28 July: Gemini can generate and edit images, diagrams and infographics inside a Doc, using the document's own text as context, and restyle several of them at once from a plain instruction. Web only, and limited to Business Standard and Plus, Enterprise Standard and Plus, and the AI Pro and Ultra consumer plans. Small thing, but method statements and site reports are exactly the documents nobody wants to redraw.
Google Workspace Updates - 06
Construction sits at 13% AI use — the lowest of any UK sector
The ONS release of 20 July puts self-reported AI use across UK businesses with 10+ employees at around 35%, up from roughly 12% in late 2023. Construction reports 13%, against nearly three-fifths in information and communications. The other number worth reading: adopters use about 1.6 AI technologies on average, so even the firms that have started have barely started.
Office for National Statistics
Read the Unit 42 write-up with your own server in mind, not a bank's. The interesting part is not that a model wrote the exploits — it is the target list. Langflow, n8n, a Marimo notebook: the sort of thing that gets stood up on a spare machine so quotes stop getting lost, then runs untouched for a year because it works. Nobody puts it on a patch schedule. Nobody remembers it has a public IP. And the economics of scanning it just changed, because finding 460 candidates used to be a week of somebody's attention and is now a paragraph of instruction.
That does not make automation a bad idea. It makes an unpatched, publicly reachable automation box a bad idea, which is a different and much easier problem. Put the thing behind a login, keep it off the open internet, update it when it asks. Ten minutes.
The awkward bit is the pattern underneath. Attackers got the compounding benefit of agents about eighteen months before the average contractor did — because they had no procurement process, no insurer to ask, and nothing to lose by trying. Same tools, same week, very different adoption curve. The ONS number says 13% of construction firms have started; the attackers are at 100%.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.