luke

security researcher. 17.

[luke @lukefr09]

apple has been exceptionally quick today. triaged 4 of these from reviewing to in progress in the past 12 hours

[luke @lukefr09]

i just hit safeguards downgrading me from fable 5 -> opus 5 -> opus 4.8 -> sonnet 4.6 -> HAIKU 4.5??

[luke @lukefr09]

anyone know how to fix this?

[luke @lukefr09]

im dying at these opus 5 commit names "fix(ui): right-align the columns that opted out" like what are we doing

[luke @lukefr09]

@D0rkerDevil same thing just happened to me with 64783

[luke @lukefr09]

@thedawgyg @slinafirinne @Nord0x yes sir! 🫡

[luke @lukefr09]

love to see it!

[luke @lukefr09]

@thedawgyg @slinafirinne @Nord0x thanks man. i hope it works out for you

[luke @lukefr09]

@thedawgyg @slinafirinne @Nord0x you're genuinely so insane. i aspire to be like you. dont wanna bug you but do you have any tips for hunting?

[luke @lukefr09]

@thedawgyg @slinafirinne @Nord0x well shoot. we will see I guess. thanks man

[luke @lukefr09]

@slinafirinne @Nord0x oh seriously! thanks, ill ask them!

[luke @lukefr09]

@slinafirinne @Nord0x this one I captured the 5k target flag! so im hoping they pay that out

[luke @lukefr09]

aghhh. still have ~15 submissions in flight so hopefully im in the next release

[luke @lukefr09]

love to see it <3

[luke @lukefr09]

@speedyfriend433 @BitcoinComfy I see you reported a couple DoSes, would you mind if I dmed you

[luke @lukefr09]

@JasonBotterill @flowersslop opus 4.7 knows with web search off. don't think this is a reliable test

[luke @lukefr09]

this is just like the early days of ai in vulnerability research. hope the researchers can handle an excess volume of emails. now on the bright side the models are quite smart so. itll be an interesting couple months

[Adam Holter @AdamHoltererer]

Good morning, I just had GPT 5.6 Pro solve an open math conjecture:

For every fixed j and k≥3,
FO_{j,k}(n)/FD_{j,k}(n) → 0,

so FD_{j,k}(n) > FO_{j,k}(n) eventually.

First prompt:
“Pick a real open mathematical conjecture you can solve. Do the breakthrough and finish it.”

38 minutes later, it returned a proof of Conjecture 1.2 from Gray–Payne–Swisher–Watson’s 2026 paper on fixed-perimeter partitions, and claimed a stronger result.

Then:
“Try to kill this proof. If you can’t, make it airtight.”
It found a gap in its own dominant-pole argument, repaired it, rebuilt the proof bundle, and ran 7.3M+ exact checks. Everything passed.

I have no clue if this is actually correct.

Math people: please try to break it:
https://t.co/QlhgZLVKy8

[luke @lukefr09]

there's not even that many hours in a week eh?

[luke @lukefr09]

@labomen001 hahaha that’s great. nothings private anymore. fwiw i did end up emailing CE so. me 1 gpt.. 99

[luke @lukefr09]

unfortunately relatable

[Air Katakana @airkatakana]

thanks to gpt-5.6, i am now a crackpot emailing quantum computing researchers at 4am

[luke @lukefr09]

@labomen001 yes… that’s slightly concerning. should have figured it can do whatever though. what gave it away?