Daily Briefing
31 October 2026 · 5 stories
🔥 Top story
1Google opens to everyone the tool that drives your screen for you
Picture an assistant that looks at your computer screen, decides where to click and types the text in your place. Google has now opened to the public the tool that makes this possible: it proposes a click, a keystroke or a scroll, on a browser, a phone or a computer. One caveat though: the tool moves nothing on its own. The developer's application has to carry out the command, in a separate space, and send back a new screen shot. Google advises against using it for sensitive decisions (payments, purchases, regulated goods) and adds a detector for orders hidden in the pages you visit. Source: Google.
2A fake Claude Code extension spent eleven hours opening a back door
Picture the official extension of a coding assistant - the one you install without thinking - being swapped for a booby-trapped version for a single night. That is exactly what happened in Cursor's extension store on October 29: for about eleven hours, two forged versions slipped in. They installed an attacker-controlled SSH key, tried to open a tunnel back to his machine, and re-launched themselves every time your computer started. Cursor estimates up to 5,000 machines were affected, and makes one point rare in the software industry: deleting the extension is not enough, because scripts stay on your system. Source: Cursor.
3Alibaba Cloud publishes its enterprise agent handbook, and its survey is blunt
Picture a cookbook written by a chef who starts by telling you that only one cook in ten succeeds. Alibaba Cloud has just published an enterprise agent handbook in the open, under the Apache 2.0 licence: thirty chapters, eighteen real cases. The most useful part is not the handbook but the survey it contains: of the teams questioned, only 18 % make it into production, and 10 % reach a 90 % success rate. Three symptoms keep coming back: harmful actions, losing the thread on long tasks, and a bill that keeps climbing. The message: the model is not the blocker, the engineering around it is. Source: Alibaba Cloud.
4Eleven publishers join forces to govern agents, with an emergency switch
Picture the control panel of a building: each floor has its own rules, but a single button cuts power to the whole place in case of fire. ServiceNow has brought eleven publishers together - Anthropic, AWS, Cisco, NVIDIA, Okta, OpenAI, Stripe and others - to govern company agents from a single platform. The order they settle on is explicit: first the agent's identity, then its isolated execution, then the emergency cut-off, and only then observation. Okta is working precisely on that third step: a switch that takes away an overreaching agent's access to company data. One caveat: this is a commercial programme, not a ratified standard. Source: ServiceNow.
5Casap raises 56 million for bank agents, funded by those who pay for the damage
Picture an insurer that does not settle the damage but finances whoever prevents it. Casap, a New York startup selling agents to banks, has raised 56 million dollars, led by Base10 Partners. And look who is joining the shareholders: Experian and Synchrony, two companies that lose money every time a card is wrongly disputed. The platform serves more than a hundred banks and is extending its agents from card disputes to debt collection. The angle matters more than the amount: agents are now funded by the ones who suffer from the problem. Source: Casap.
📡 To watch
Zest books your table by text message, and the agent becomes the product
A restaurant concierge that no longer speaks through an app: you send it a message, it suggests places based on your real habits, checks availability and books through OpenTable, Resy or SevenRooms. In a group, it mixes everyone's tastes and settles it by vote. The mobile app was dropped in favour of the conversation. Source: TechCrunch.
GitHub now scans your pull requests for agent traps
GitHub is adding detection of agent-specific flaws to its automatic code analysis: an order hidden in a web page or a file, a leak of sensitive data, an agent that does too much. The analysis runs on every merge request and suggests a fix. The risk experts cite most often is the first one: the instruction hidden in a page the agent visits. Source: GitHub.
Island wants an enterprise browser built for agents, not for humans
Today an agent goes through an ordinary browser with an extension, so it can read everything and do everything: no layer tells the human apart from the machine. Island, which sells a secured browser to large companies, wants the opposite. In its version for agents, the agent inherits its user's identity, every tool is allowed application by application, and each action is attributed to the agent or to the person. Source: Island.
Sema4.ai takes repetitive work out of agents to cut the bill
The Sema4.ai platform moves repetitive work out of the models: queries and calculations validated once are reused instead of being asked of the model every time. The company reports up to 80 % fewer billed tokens and runs up to four times faster. Same logic: the agent's memory keeps each customer's rules and applies them on every run. Source: Sema4.ai.
📊 Trend
Today, control is changing hands. Google opens to everyone the tool that makes an agent click for you, but points out that the application does the moving - and that payments must stay blocked. Alibaba measures that only 18 % of teams make it into production. ServiceNow gathers eleven publishers to put the agent's identity, its isolation and its kill switch on the same level. GitHub, for its part, puts trap detection into the automatic review of every code change. The thread running through it all: the era when you delegated everything to the agent and checked afterwards is over. You check first, you limit first, and you can still cut everything.