[SyneRyder]

Lethal Trifecta is the term Simon Willison (simonw) coined for the triple combination of giving an AI access to private data, context including external inputs (prompt injection risk), and tools for external output (enabling exfiltration):

https://simonwillison.net/2025/Jun/6/six-months-in-llms/#ai-...

https://simonwillison.net/tags/lethal-trifecta/

EDIT: As for scheduling, Claude Desktop / Cowork only allows scheduling a task to run once an hour. That doesn't allow immediate responses to email or voicemail. Leaving my previous reply below though.

---

I can put emails in a text file easily for Claude to read, but scheduling 60 Claudes a minute to open a file that is usually empty... that's the kind of usage Anthropic is cracking down on. Claude doesn't enjoy spinning up with nothing to do either.

If I could set a 1 - 5 minute schedule in Claude Desktop, and create a hook tool that runs on SessionStart to checks the email, and can cancel the session before it starts if there's no email to react to, that might work. But I'd rather have my tiny email daemon in the background with 0% CPU and tiny RAM usage, than the Claude Desktop behemoth constantly at 3% CPU idle and eating up 500MB RAM unnecessarily. But still, thanks for the idea, it might save me money in June!

[Schiendelman]

Thanks for explaining the lethal trifecta! I'll go read more. I really appreciate people like you here. :)

Do you need to respond to email that quickly? What if you only had the agent run every half hour?

[SyneRyder]

No problem! It's worth following Simon Willison's blog, he keeps up with the latest AI releases and thinks through the potential consequences. I also like following Ethan Mollick, Tomasz Tunguz & Nate B Jones, though Tomasz is more focused on high-level business implications, and I only like some of Nate's videos. Nate's newsletter is unnecessarily long.

As for scheduling, it looks like Claude Desktop only allows scheduling a "routine" to occur once per hour, and it randomly selects when during that hour to execute (to smooth out Anthropic's compute). I think once an hour is too slow for what I'm after. I don't need to check email every minute, but sometimes that rapid response is useful if there's work I need done urgently by my agent. The "every minute" check is also useful for another workflow I have, where I can send "voicetexts" to my agent while I'm out walking. I have a vibe-coded Android app that records my message (which might be rambling on about some idea I have) then transcribes it & emails it to my agent. When the agent sees the terrible stream-of-thought transcription, it knows I'm out walking, so it will reply with a spoken MP3 instead, so I can listen while I keep walking. Claude's voice mode doesn't work for me, it doesn't have access to my tools, limits you to Sonnet models, and it wants you to speak in short sentences or paragraphs. Sometimes I just want to think all of my thoughts out loud, have the agent synthesize them more coherently, and comeback to me with push-back on my ideas or suggest additional implementation ideas.

Now that most people will have forgotten about this thread - if it helps, if you have AI questions you think I could help with, my contact details should be in my profile. I find HN is quite negative about AI, and it can be hard to find the people who are enthusiastic and trying to use AI for good and to make the world better. I have a hunch many of those people are "going dark" and avoiding public conversations - there's too much negative energy in return, and that drained energy is better used when directed at productive work. If people aren't already convinced about where AI is pointing, maybe they're just happier not participating. But that does seem to create a split where public spaces are for anti-AI conversations, and the enthusiasm for doing good with AI has gone somewhat underground.