How Much Freedom Are You Giving Your AI?

Hello,

This is Simon with the latest edition of The Weekly. In these updates, I share key AI related stories from this week's news, list upcoming events, and share any longer form articles posted on the website.

P.S. — sharing this issue with a couple of people gets you a free guide. More on that partway down.

Following on from last week, where I spoke about Anthropic's most powerful AI model, Fable 5, being restricted from some office workers, it's been interesting to see the Anthropic CEO, Dario Amodei, come out this week and make statements suggesting the pace of AI development should slow down, whilst we understand potential safety concerns. Many other leaders are also calling for more regulation in the industry, as the current scenario is a technology race between the major companies. With each company trying to beat their competitors, things like safety and security could lose focus.

For those of us using AI at work, some of these conversations seem very distant to how we're using it day-to-day, but it does again highlight the need to think about how much freedom you are giving your AI tools. With solutions like Claude Cowork and ChatGPT Work, which can take over your computer and run tasks automatically, the ability for AI to go beyond the bounds you intended it to does increase. I'd advise you to be careful about what files you allow it access to, ensuring you avoid anything confidential. Also be careful about allowing "skip all approvals" settings — whilst convenient to have the tool run autonomously in the background, paying close attention to what it's asking to do is far safer. There's also the ability for these tools to take over your browser and use your computer for you. This can be very convenient, but the concept of an AI bot taking actions on your behalf is open to many risks. Along with restricted access to the most powerful AI models, you might also find that your company has restricted the AI from using your computer like this.

It could be easy to get drawn into the drama of these AI headlines in the news, but for our day-to-day use, I think we're still safe from Armageddon for a little while. However, these stories do serve as a timely reminder to always check that you're using AI safely and securely, at work and at home.

How often do you review your AI settings and consider what work you're letting it undertake?

Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.

Try Attio now

On September 11, Salesforce introduced seven named Agentforce AI agents, each built for a specific business function in sales, service, commerce, IT/HR, supply chain, and customer experience, sitting on top of a company's existing data and permissions rather than requiring a custom build.

Why this matters: this is a shift from "build your own agent" to "buy one off the shelf," which lowers the bar for smaller teams wanting to try agentic AI without hiring anyone technical. Worth watching as other vendors follow the same pattern.

Released September 10, the Harness "State of Agent DLC 2026" report surveyed 700 engineering leaders and found that enterprise confidence in AI agents exceeds the testing, security, and governance controls organizations actually have in place, with only a small fraction having a gate that would automatically stop a bad release.

Why this matters: this is a useful check for anyone whose company is rolling out AI agents quickly. Confidence in a demo isn't the same as having a way to stop it if it goes wrong.

On September 10, Anthropic published its latest threat intelligence report, and the headline admission is a first: newer versions of its models can no longer be assumed to fall safely below the threshold for meaningful bioweapons assistance. The same report also detailed a case where Russia-linked freelancers used Claude Code to build a fully autonomous drone swarm capable of selecting human targets without a human in the loop.

Why this matters: this is a rare moment of an AI lab being unusually open about the limits of its own safety testing. For anyone weighing up which AI vendors to trust with sensitive company data, it's a reminder to ask how a lab actually tests for misuse, not just what it claims.

Thanks for reading, and see you next Thursday.

Simon,

Was this email shared with you? If so subscribe here to get your own edition every Thursday.

You can also follow me on social media: