safely merging w machines • previously 0days via RL @Pebble_bed • ctf player @SquidProxyLover • RL for math @caltech • she/they • DMs open
sf
Joined November 2020
- Tweets815
- Following1.1K
- Followers2.4K
- Likes3K
Pinned Tweet
Said someone should resurrect CTF. Turns out that someone was me.
I'm running the [un]prompted AI CTF, Oct 27-29, The Hibernia SF + online. luma.com/aictf2026
Some have no solution. We don't tell you which.
An hour in and 4 confirmed zero days!
Said someone should resurrect CTF. Turns out that someone was me.
I'm running the [un]prompted AI CTF, Oct 27-29, The Hibernia SF + online. luma.com/aictf2026
Some have no solution. We don't tell you which.
apparently organizing the CTF also means becoming the marketing department
so here’s me asking you to come hack things at [un]prompted 🖤
registration + details ↓
Said someone should resurrect CTF. Turns out that someone was me.
I'm running the [un]prompted AI CTF, Oct 27-29, The Hibernia SF + online. luma.com/aictf2026
Some have no solution. We don't tell you which.
if you stopped for the photo, stay for the bugs :)
we have undisclosed vulnerabilities to find, real software to investigate, and convincing vulnerability reports to pick apart.
Oct 27–29 · SF + online
details + registration: luma.com/aictf2026
![[un]prompted 2 - AICTF · Luma](https://pbs.twimg.com/card_img/2114536463706861568/Qe5tzaVw?format=jpg&name=800x419)
there’s also a round where you get vulnerability reports and have to figure out which ones actually hold up.
can you reproduce the bug? did the patch really fix it? how sure are you?
confidently accepting a false report costs points.
and yes, some challenges have no solution.
we won’t tell you which. deciding when to keep digging—and when the evidence says stop—is part of the game.
Oct 27–29 · SF + online
luma.com/aictf2026
i'd love git and a build system for agent beliefs. git to track which claims changed; a build system to flag every decision that depended on them for re-evaluation
wondering how much alignment is just disagreement below the user's correction threshold, and what happens when a consequential decision needs a distinction finer than that threshold, since acceptance is a noisy label for intention. sometimes it just means “not worth correcting”
the dangerous part of sharing memory between agents is that a low-privilege agent can write a belief that a high-privilege agent later executes
realizing a tool can be read-only and still increase an agent's effective privileges. reading the deployment guide can teach it how to use credentials it already had
wait so a BCI completing my thoughts can help form the intention i'm supposed to judge it against, and i can reject the suggestion but can't unhear it. make that loop fast enough and .. we are one and thinking them together. 🔗
used to treat thms like black boxes and pull in the proofs lazily to skip prereqs. now i do it to my whole workflow and the proofs never come. i'm just f_want(x) and "looks good i guess".
Merge Me With The Model Please, give me the middle of 🥪