The big problem with that is every election you have to hope people don't vote in a new government that's keen to abuse the systems that were installed previously
@simonw
-
CaMeL Pattern: Building More Sophisticated AI Systems
By
–
The most convincing pattern I've see for building more sophisticated things than that is CaMeL, but it's extremely hard to implement
-
Securing AI Agents: Preventing Malicious Tool Exploitation
By
–
The easiest one is to make sure you don't expose all three legs of the lethal trifecta at the same time – and also that you design things to assume that anyone who gets malicious content into your agent can take full control any of the tools it's allowed to execute
-
Building Development Tools with AI on Mobile Devices
By
–
Yes tongue in cheek, but honestly I really do get a ton of development work done on my phone these days – probably around half of the things on my https://
tools.simonwillison.net site were built by prompting Claude or ChatGPT on my phone and pasting the result into GitHub -
Desktop Requirement for App Development Still Relevant in 2025?
By
–
"Open this page on a desktop browser to build your first app" – people still use laptops to write apps these days, in 2025!?
-
ICPC Results Show AI Problem-Solving Beyond Training Data
By
–
No, the reason the ICPC result is so interesting is that the challenges were created specifically for this year's competition and could not have appeared anywhere in the training data
-
Salesforce Security Bug: Expired Domain in CSP Header
By
–
Wrote more about the Salesforce bug on my blog – one key detail is that Salesforce were using a CSP header to prevent loading images from untrusted domains… but one of their allow-listed domains had expired!
-
OpenAI Training TFLOP Figures: Source and Transparency Questions
By
–
Where do those training TFLOP figure come from? I didn't know OpenAI shared numbers like that
-
Clarifying Prompt Injection vs Jailbreaking Misconceptions
By
–
Plus I'd had limited success getting people to understand "prompt injection" because too many people assumed it meant people "injecting bad prompts" into the system themselves, conflating it with jailbreaking
-
Lethal Trifecta: Data Exfiltration Through Prompt Injection Attacks
By
–
That's regular prompt injection. I came up with the "lethal trifecta" framing to help people understand the data exfiltration subset of prompt injection, because the same exact bug kept on showing up again and again and again across different systems