PandaSt0rm

I'm sitting on so many full chain 0days I don't actually know what to do with myself... I miss the days of spending weeks to months tinkering on a single project, now I just have 100+ agents working on 10 different code bases and my brain feels like its permanently on fire... Its total mental overload 😩

You know... Last and this weekend I emailed (they wanted disclosure via email) a bunch of fully built and validated chains and patches to various linux projects which are nearly ubiquitous across the ecosystem. Some things kinda "holy shit holy shit holy shit patch yesterday" urgency. One has a "point at ip on the internet sometimes (like a 60% hit rate) get a shell" PoC attached. Yet to hear back from anyone :(

Kind of a funny comparison. Fully used my SuperGrok Heavy quota (was my last week, reset would have been in 3 days but i'm not renewing). Only ~1.5B tokens a week on Grok 4.7 for $1300 in equivalent API costs, vs $30 for nearly 4B DeepSeek tokens lol... Pretty much the same capability. DeepSeek was much faster too

Oh? @morphllm

Who started this "AI attacks run at machine speed" nonsense?

Holy trip down memory lane...

I swear fuzzing is worse than playing the slot machines. At least you can run out of money pulling that lever. With fuzzing: you gently caress the code to get it to tell you its secrets, you build the harness, you run it, waiting with baited breath. BRIGHT RED LETTERING POPS UP, YOU HAVE A CRASH. Triage starts. Adrenaline building... Its not exploitable, low impact... fuck we go again...

I don't get the hate against @GeminiApp It's been my goto model since argon went public, I prefer it even over opus 5.5. Argon is probably the first model that does better than the benchmarks indicate. GPT Sol 6.1 I won't let near my code bases anymore, it just wreaks absolute havoc. I know this is not a universal experience so its highly likely a side effect of either the languages i'm working in or the complexity of the work and my agent setups and skills. Opus 5.5 is okay, happy to give it a prompt and then background omp and come back a few hours later and check on it. I just don't personally like how claude writes code in general so not a huge fan. Argon I'm in love with. I can give it the vaguest of bugs and it will easily resolve them. It seems to just really *get* me and how I plan and architect projects. Rarely does it need to ask 30+ minutes of questions like opus does. I've yet to find a task I can't just toss at it and it will come back with a fully functional solution.

I kinda wonder if this whole 'cyber restriction' thing going on in AI is a way to both create and then grab a pretty profitable market segment. From what I've heard companies are paying stupid amounts of money to be able to use glasswing models for eg. If the ai companies can pump up the specter of "we're all fucked because of AI" BUT with the fine print of "only if you dont secure your systems with our super cyber capable model" they're likely getting or going to be getting huge chunks of security budgets from orgs :/

It's kinda funny how argon keeps flagging *specifically* on objdump... tho in general I hope this isnt starting to shit towards a more restrictive model because it was not doing this a couple of days ago :/

Opus 5.5 Safe Guards just flagged on building an auth service in go... You cannot be serious....

This whole PHP thing has brought back some real nostalgia... I wonder if we will ever go back to the days of walking into a decrepit cafe with a laptop, a cellular dongle and a thumb drive, sitting down opposite some nameless dude in a fedora and a Hawaiian shirt, firing up the laptop, proving a 0day works on a few targets (for php hilariously enough), copying the code and docs onto the thumb drive, handing over the thumb drive to Mr Nameless, and then walking out with a duffle bag stuffed with cash; or if AI has killed off the "funner" 0day markets entirely.

This is kinda painful not gonna lie, took like 5 minutes for the old cvp program to be approved, 10 days and counting here 🥹

As much as I really like grok 4.6 I have spent the last week undoing and fixing basically everything it implemented in a project lol, the code looks fine and passes tests but holy crap its 2nd order effects broke nearly everything... I reaaaaally hope 4.7 has a "broader project horizon" for planning and implementation

If this is not the future I don't know what is... Currently have @bot overseeing model tuning experiments, it found some issues with my training data it fixed, it then fixed a logic bug in my training recipe (its a non standard approach I'm experimenting with). It ran the entire tune, then did a quick evaluation on the outputs, identified garbled outputs, suggested some updates to my maths, I tossed that into GPT Pro, it caught an issue, back into bot, finished that run, muuuuch better outputs, starting tuning knobs... We're now 14 runs deep, its keeping a ledger of the failures and successes and building comparative benchmarks. So far I've got a nearly 2x uplift in the domain I'm training on without dropping capabilities anywhere else... This is what it must feel like to be at one of the ai companies but with practically unlimited compute instead

Is opus actually unable to do work on its own anymore? Been experimenting using opus subagents to independently judge the assumptions and code-based hallucinations opus has been wildly coming up with and its been pretty successful, its just insanely slow since the planning stage now takes like an hour before any implementation actually happens; its sort of insane how "lost" opus 5 seems to actually be, its soooo hard to get it to work by just prompting it without all the "recursion" bs now

I think @bot is my favorite research assistant now. Asked it to put together a whole bunch of research for a new project last night, woke up this morning to a 400kb .md doc that's far deeper than anything I've ever prepared and it even caught some really interesting adjacent work already done by some university teams lol. The doc itself is far more usable (there is basically no fluff) than anything i've seen other companies "deep research" offerings generate.