Ori Ron and I found a cool way to attack the HITL, by convincing it to inject content and markup right after commands. Anyone would press Yes if the attackers control the question.

Interesting. Thanks. Might make sense for problems there to grow as well - both the ability to label at an increasing scale, AND the ability to separate fresh and slop as quality improves.

But how can they tell? Inevitably at the rate of things a vast majority will be slop, or a mixture of organic and slop.

But the whole prospect of curation and purity goes out the window on an internet of slop, such that even a few short years from now even multiple models with varied datasets will become inbred.

So this might be a stupid question but - if #LLMs that feed themselves ruin their own datasets, wouldn't inbreeding of a several LLMs feeding one another also inevitably result in the same? Wouldn't an LLM-heavy internet inevitably still become a sort of inbred LLM amalgamation meta-model/dataset?

We did a thing with Claude Code checkmarx.com/zero-post/by...
It's very cool, but using it to gate automated security code reviews is very, very dangerous

Want to see a free, #OpenSource, developer-friendly tool for preventing secrets leaks? Checkmarx Zero's Tal Folkman will be on-site BlackHat #Arsenal (#BHUSA) to demo Too Many Secrets (2MS), available from buff.ly/Yng76l5

Mark your calendar! 2pm (local time) on 6th August, at Arsenal Station 5

I just can't believe the ysoserial.net exploits for BinaryFormatter still fucking work tbh. TypeConfuseDelegate payload as a ViewState via the MachineKey. This shit is old school.

It's no surprise at still finding decade old gremlins inside MS code, but why are these endpoints are still exposed?

wow, $13 billion in earnings? I'm just guessing here but that sounds to me like at least $20b in revenue. And therefore at the moment... about at least $20b short.

#CVE-2025-6514 being 9.6 is a bit bullshit. It's kinda cool, but I'm sorry - if you connect with #MCP, a protocol for running code, to an untrusted MCP server, and it runs code - it's a footgun, not an RCE. The exploit itself is very interesting though.

#TACO? Surely you mean Wrap of America.

My condolences to the United States for having lost the Cold War.

Protectionism always works out so well. Also very capitalistic, much free.

This has always been true. Space magic used to be just another mysterious layer of what is essentially a lush and rundown universe that is old and full of junk and lasers. The prequels made Jedi front and center, which is why Andor and Skeleton Crew are the best Star Wars since orig trig IMO.

If by masculine he means the really old-fashioned kind then well Facebook is well on its way to dying bitter and alone

Oh then you're going to love 100 hours of minigames and Ubisoft-grade exploration. Still not as bad as how they cripple the original narrative's momentum at every opportunity though, mostly to introduce weird childish meta fanfic. Maybe less would have been more.

-ish. I think the general new GoW mechanisms are just not that great when compared to the whole Dark Souls ripoff space, it's a very mid game mechanically. But exploration feels less on-rails. I do miss the old gameplay because it felt tighter.

And finally part 4 of 4 of my @hf.co security research - bypassing poorly designed model scanners multiple times to smuggle malicious code because the detection mechanism pattern rhymes with "chitty glocklist" 🤷🏼
checkmarx.com/blog/free-hu...
#appsec

Part 3 of 4 of the Hugging Face security research - so many model-loading HF integrated libs are vulnerable to code execution, but it's all an open secret that no one will fix
checkmarx.com/blog/free-hu...

plz delete this is literally the plot to Scream XXIV

Aaaand here's part 2 of my horrible Hugging Face blog, it's a deeper dive into code execution in protocols used by open source LLM frameworks
checkmarx.com/blog/free-hu...

So I found a ton of interesting crap on Hugging Face, some of it known and some of it less so. Started a 4 part blog about it.
Here's part 1, about configuration and excessive trust in ReadMe files
checkmarx.com/blog/free-hu...

How fortunate, I closed my Twitter account two weeks ago and hey what do we have here? Hope this one catches, Apartheid Karen can eat ass.