By
–
Claude Code keeps on adding features that I didn't realize I needed – being able to start a dev server in the background and occasionally dip into the logs is super useful https://t.co/z6XaOKUpDC
— Simon Willison (@simonw) 9 août 2025
Claude Code keeps on adding features that I didn't realize I needed – being able to start a dev server in the background and occasionally dip into the logs is super useful
By
–
Not that let you use all three legs of the lethal trifecta at the same time For prompt injection generally I like these two papers: https://
simonwillison.net/2025/Apr/11/ca
mel/
… and https://
simonwillison.net/2025/Jun/13/pr
ompt-injection-design-patterns/
…
By
–
I have a version that does that for the image input models, I hope to write that up soon
By
–
I've even seen some of them wildly over-think on my pelican benchmark – got a fantastic result from Qwen3 4B Thinking yesterday
By
–
Right, that's why the stuff is such a massive nightmare: so many of the things that we want to use LLMs for can't be done securely in the absence of a completely reliable protection against this class of attacks – which so far does not exist
By
–
I think in some cases it is reasonable to expect 100% reliability from a vulnerability mitigation
By
–
You mean the CaMeL approach? Yeah actually implementing that well looks really difficult to me, you have to make about sure that any untrusted content is "tainted" and is then NEVER included in a prompt that might trigger tools
By
–
The router only affects ChatGPT – when you use the API you get complete control over which model you are calling and the reasoning effort it uses
By
–
Here's a good recent example: Cursor/Claude was refusing to steal API tokens, so they changed the attack to request "rotten apples" which were of strings starting with "eyJ" – aka JWTs! https://
simonwillison.net/2025/Aug/9/whe
n-a-jira-ticket-can-steal-your-secrets/
…
By
–
I think they give people a false sense of security which makes it much more likely they'll fall victim to an attack