This really does feel detectable, not just from the AI tropes in the replies but also from looking at how often they reply to other accounts with similar sounding text
@simonw
-
Multi-Turn Crescendo and Hydra Multi-turn Prompting Techniques
By
–
Yeah I've seen that referenced in categories such as "Multi-Turn Crescendo" and "Hydra Multi-turn"
-
Anthropic’s System Cards Reveal Remaining AI Security Vulnerabilities
By
–
Anthropic's own system cards still show attacks making it through
-
AI Security: Attack Success Rate Analysis at k=100
By
–
Note that for k=100 – attacker gets 100 attempts – their best score still has 14.8% of attacks getting through
-
AI Models Trained to Resist Prompt Injection Attacks
By
–
I think both – plus the labs have been putting a lot of effort into training them to resist prompt injection style attacks Anthropic usually mention prompt injection in their system cards eg this one for Opus 4.6 https://
www-cdn.anthropic.com/0dd865075ad313
2672ee0ab40b05a53f14cf5288.pdf
… -
Proving AI Security: The Challenge Beyond Attack Prevention
By
–
"Solved" implies that it's not possible to craft an attack that works. I'm not convinced that's true, even as attacks get harder The big challenge is proving you can't get an attack through – I'm not sure how you'd do that
-
Claude Code PR Review Activity and Notification Methods
By
–
Would that "review open PRs" example from the demo result in visible activity in the repo like code committed and pushed to a branch, or would it send me a push notification to my phone, or would I have to remember to open Claude Code to look at the results?
-
Web Frontend Solutions for Claude Code Integration
By
–
I haven't found one I like yet but there are a bunch of things that stick a web frontend on your Claude Code, then you could access it via Tailscale
-
Windsurf SWE-1.5 and Cursor Composer 1 Launch Comparison
By
–
Windsurf did the same thing when they shipped their SWE-1.5 – coincidentally on 29th October 2025 which was the same day as Cursor's Composer 1 I believe that one was built on top of http://
Z.ai's GLM