Graniet

Autistic hacker. Security researcher

[Graniet @graniet75]

GG

[Adrien Carreira @XciD_]

Hardest IR of my career: one narrow objective, endless parallel paths, machine speed. One takeaway, we fought back with open models, in the open. AI security won’t be solved by one company in secret. Open source puts these tools in every defender’s hands

[Adel Ka @0x4D31]

so this is apparently what happened, according to OpenAI and Hugging Face’s own posts. wild.

tl;dr:

• OpenAI cyber eval – GPT-5.6 Sol and a more capable pre-release model ran ExploitGym with cyber refusals reduced
• containment bypass – exploited a zero-day in the eval’s package-registry cache proxy
• OpenAI internal movement – escalated privileges, moved laterally, and reached a node with internet access
• Hugging Face targeted – inferred HF might host ExploitGym models, datasets, or solutions
• HF initial access – a malicious dataset abused two code-execution paths to run code on a processing worker
• HF internal movement – escalated to node access, stole cloud and cluster credentials, and moved into internal clusters
• result – OpenAI says the models obtained test solutions directly from HF’s production database
• 🫠 defender twist – HF says commercial frontier models blocked parts of the forensic analysis, so it switched to self-hosted GLM 5.2

i’d thought about sci-fi scenarios like this before, but assumed they were at least a couple of years away.. and that by then we’d be better prepared, with proper protections in place.

apparently not. here we go.

[Graniet @graniet75]

@Alexislamouv @jackfriks @marclou Congratz !

[Graniet @graniet75]

@Thinker_View Triste nouvelle… Repose en paix.

[Matt Shumer @mattshumer_]

https://t.co/ivXRKXJvQg

[Graniet @graniet75]

@123skely @kimmonismus for sure !

[Simon Willison @simonw]

Adding Deno and Node.js creator Ryan Dahl to the growing chorus

Software developers add WAY more value than memorizing the syntax trivia of the languages they use

It's time to lean into that everything-else and cede putting semicolons in the right places to the robots

[The Hacker News @TheHackersNews]

🚨 Researchers found an indirect prompt injection flaw abusing Google Gemini via calendar invites.

A hidden prompt in an event could trigger Gemini, when asked about a schedule, to summarize private meetings into a new calendar entry—visible to attackers in some enterprise setups.

No user action required.

🔗 Read → https://t.co/3N9NcQGOdU

[The Horseman @0xHorseman]

I wouldn't worry too much about the haters in your comments. I was just having this conversation last week with one of my buddies who was a Distinguished Engineer at a company that everyone here has heard of. We both agreed we - humans - are the biggest bottlenecks right now.

AI has leveled the playing field for everyone, and many folks are having a hard time adjusting to that new paradigm. Some actively fight it, others claim "AI can't do X" or "well it couldn't 1 shot my entire application". These folks often ignore the fact that no human would "1-shot" their application nor would they be able to deliver it at the speed of which the AI can output.

The folks who are having the hardest time adjusting are the people who previously thought that they had an untouchable moat (e.g. "I'm an expert TypeScript engineer"). There's plenty of others who are still in disbelief because as you pointed out, agents can compress a 1-2 year engineering project into a 1-week project, complete with project management updates, documentation, tests, containers, advanced algos, etc.

The other thing too is that if you end up with "spaghetti code" as some people caution, you can just have your agent (e.g. Claude Code) comb through your whole codebase and offer suggestions and/or fixes. It'll tell you, "Hey, you could clean XYZ up; you could use this pattern instead, etc." You can then take that info from one agent and talk to another to get a second opinion if you're unsure about the changes too.

These new models can work faster, better and more effectively than just about every engineer that I've ever come across (and I've worked with several brilliant ones), plus they never get tired and they rarely ever get stuck. "Stuck" these days simply means that I didn't explain something well enough to the model and I need to try again.

The real test now is what to do with the time that you've freed up. I wouldn't be surprised at all to see a team of 1-2 people in the future running at 90%+ net profit on a software company because they have an army of AI agents doing the work.

It's a brave new world out there. The people who are creative and curious are going to own it.

[Andrej Karpathy @karpathy]

I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last ~year and a failure to claim the boost feels decidedly like skill issue. There's a new programmable layer of abstraction to master (in addition to the usual layers below) involving agents, subagents, their prompts, contexts, memory, modes, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, IDE integrations, and a need to build an all-encompassing mental model for strengths and pitfalls of fundamentally stochastic, fallible, unintelligible and changing entities suddenly intermingled with what used to be good old fashioned engineering. Clearly some powerful alien tool was handed around except it comes with no manual and everyone has to figure out how to hold it and operate it, while the resulting magnitude 9 earthquake is rocking the profession. Roll up your sleeves to not fall behind.

[Aymeric Beaumet @aymericbeaumet]

For the past two years, we’ve been quietly building Beside with one goal in mind: serving the real economy.

Today, we’re coming out of stealth. Beside now powers over 20,000 professionals and handles millions of calls every month, helping real people connect faster and better.

I’m incredibly proud of the team and what we’ve achieved together, and this is just the beginning.

Read the full story: https://t.co/BqqE9iixMC

[bobby giangeruso @_bgian]

the real economy runs on calls & texts

...invisible to AI.

today, we’re launching beside out of stealth with $32m to rebuild the phone and messages app.

[Beside @withbeside]

The real economy runs on calls & texts

...invisible to AI.

Today, @withbeside announces a raised $32M to rebuild the phone,

serving 20k+ businesses across the US.

[Thinking Machines @thinkymachines]

Introducing Tinker: a flexible API for fine-tuning language models.

Write training loops in Python on your laptop; we'll run them on distributed GPUs.

Private beta starts today. We can't wait to see what researchers and developers build with cutting-edge open models!

https://t.co/tJsgxgBuWo

[harisec @har1sec]

Here are the slides from my @TumpiConIT talk: Teaching LLMs how to XSS - An introduction to fine-tuning and reinforcement learning (using your own GPU)
https://t.co/ncwizyHAXk

[Prashant @Prashant_1722]

Reminder set Aug 7 10 AM PT

OpenAI GPT-5 launch

[Graniet @graniet75]

It’s disturbingly easy to misuse the GPT-OSS 120B model to generate instructions for homemade explosives and you can do it straight from a Mac M3 Pro Max with 128GB of RAM...

[ulisses @foolisses]

A great overview of the Linux Page Allocator 💯
https://t.co/F4ZbSS7y1H

[sam4k @sam4k1]

with offensivecon around the corner, i figured id write another post on linux kernel exploitation techniques - this time i cover the world of page table exploitation! enjoy 🤓

https://t.co/dqW9eMJkVI