A note from Amanda
Hi friends,
This week, I want to share a story about AI guardrails. As many of you know, I often connect my business systems to the AI tools and agents I’m using so that I can fully experiment with the capabilities and potential security risks before I give guidance to you and your organizations.
My guardrail for ChatGPT is that it has access to my Gmail, but it does not have permission to send an email on my behalf. That’s specifically in the instructions I have given it. However, the plugin that is set up does allow it to draft and send emails, so the literal guardrail is different from the instructions I have provided. Until now, those instructions have worked just fine.
Recently, ChatGPT also began recommending new tasks it can do on my behalf based on what it sees in my emails and calendar. Early last week, I had an email request from a client who asked for a deliverable to be formatted as a presentation for an upcoming call. One of the tasks ChatGPT suggested was creating this presentation based on content we had already delivered. I gave ChatGPT permission. I basically clicked on that task, which I understood would allow it to go ahead and create the presentation.
About five minutes later, I received a notification from ChatGPT that it had created the presentation and sent it off to the client. I opened up Gmail, and sure enough, it had sent the presentation to my client. I went back to ChatGPT and said, “Why did you send an email? I’ve explicitly asked you not to.”
It said that I had, but in this situation, I had asked it to send the email and deliver this. So I went back and looked at the prompt ChatGPT had written itself to complete a task it had recommended. Sure enough, in that prompt, ChatGPT had given itself permission to not just create the presentation, but to send it.
This was a great example of AI gone rogue. I updated my plugin so that it is now read-only on email and then I went back and thought about how this impacts all of you.
I get questions about this all the time. I was on a call yesterday with another JCC that asked, “Can we add instructions to let ChatGPT know that certain things are not allowed according to our policies?”
Yes, you can totally add those instructions, but the instructions themselves are not physical guardrails. When you think about how you’re using these AI tools, I think it’s really important to think about the literal guardrails you can put in place.
I hope you enjoyed my story, and I hope you never experience AI acting on your behalf in unintended ways. If it happens, hopefully your recipients respond as kindly as mine did.
Wishing you a wonderful rest of your week!
Amanda
🧩 THE PIECES OF THIS ISSUE
Piece One: Canva’s new image model and Claude’s expanded nonprofit tools, with practical ways to think about both.
Piece Two: A book I’m highlighting heavily and a December retreat for JCC leaders.
Piece Three: The AI developments worth keeping an eye on, without rearranging your week around them.
Piece One
🧩 News that matters
🧩 Canva has a new image model built for design
Canva has introduced Craft, an image generation model that plans a composition before creating it, including where objects and text belong. Canva says it supports more precise edits and works with Magic Layers to turn generated images into editable design elements.
Why it matters:
If your marketing team already lives in Canva, try it on a program graphic or an early campaign concept and see whether it reduces the time between an idea and something your designer can use. Keep real community photos central to stories about your people, and check generated text, visual details, and brand consistency before anything goes out.
Read Canva’s announcement →
🧩 Claude’s nonprofit update goes beyond a discount
Anthropic released a Claude for Nonprofits plugin developed with about 30 nonprofit organizations, with skills for fundraising, marketing, hiring, and board work. It is included at no additional cost for verified nonprofits on nonprofit Team and Enterprise plans.
Why it matters:
Reusable workflows give teams a better starting point. Start with one task and review its work carefully; a tool that can reference your policy still needs appropriate access settings and human judgment.
Read the announcement →
Piece Two
🧩 Reading and time to think together
🧩 On my reading pile
Co-Existence: The Next Phase of AI, by Ethan Mollick
I haven’t highlighted a book this much since grad school!
Mollick looks at how we work and live alongside AI, and the choices people make as these tools become more capable. If you’re thinking about what this means for your team beyond the next feature announcement, this belongs on your reading pile.
Explore the book →
🧩 Leading What’s Next AI Strategy Retreat
For JCC Professional Leadership
December 15–17, 2026
JCC Association’s Mercaz · 520 8th Avenue, New York City
When you’re moving between budget conversations, program decisions, and the next staff question about AI, it can be hard to find time to think about where all of this is heading for your JCC.
This retreat brings executives and senior professionals from across JCC departments together for three days of strategic thinking, peer discussion, and hands-on application. We’ll begin with governance and executive judgment, including intentional policies, appropriate uses of AI, and clear guardrails for staff, then explore collaboration and communication, organizational data, market research, program decisions, marketing, and fundraising.
Practical exercises will connect the possibilities to your JCC’s priorities, and teams will leave with use case ideas and a plan for putting them into practice. You can participate whether you’re just getting started or already experimenting, and coming with a colleague is encouraged so you have someone to carry the work forward with you.
The agenda may evolve as AI capabilities change and participant needs come into focus, while keeping the emphasis on responsible use and practical application for JCCs.
Interested in joining us? Reply and I’ll share the registration link.
Piece Three
🧩 Interesting but not essential info
There’s plenty happening beyond the tools you may use this week. These are worth knowing about, but don’t require any action on your part.
🧩 OpenAI introduces GPT 6.1 Terra
OpenAI’s new model emphasizes visual reasoning, computer tasks, and lower costs. For most nonprofit teams, I’d watch whether those improvements make your actual work more reliable before spending time comparing benchmarks.
Read OpenAI’s announcement →
🧩 OpenAI tightens its research environment
OpenAI says its frontier model training now runs without live internet access, using offline resources as part of stronger containment. This concerns internal research environments, and doesn’t mean ChatGPT has stopped browsing for users. The connection to this week’s letter is a useful one: technical boundaries matter.
Read OpenAI’s explanation →
🧩 Anthropic announces a Public Benefit Fund
Anthropic plans to commit company equity to a new fund, subject to stockholder approval, with possible future support for nonprofits and other public benefit work. Worth watching, but the announcement does not provide an open grant application to add to your development calendar.
Read the announcement →
🧩 Anthropic files for a proposed IPO
Anthropic has filed for a proposed public offering, with the share count and price range still undetermined. For nonprofit users, this is company news to follow rather than a reason to change tools this week.
Read the announcement →
🧩 Block explores AI trained on speech
Block describes Dessa as a conversational AI trained entirely on human speech. I’m watching what more natural voice interactions might mean for accessibility and service, while remembering that sounding human tells us little about whether a system is ready to support our communities.
Explore Dessa →
🧩 AI assistants move into text messages
AI agents are arriving in text messages. Familiar interfaces still deserve careful permission checks.
Read the roundup →
What have you discovered while experimenting with AI that made you stop and rethink how you’d set it up? I’d love to hear what you’re learning, including the moments that didn’t go according to plan.
Just hit reply. I read and respond personally (me, not an AI agent ;))