This is insane. Open source models will reach the coding level of Mythos within 6 months and will be able to compromise any server that isn’t patched between now and then. We are entering uncharted waters.

What else are they going to patch? What other things does those companies get? What else is this Claude good at? Are they blocking all aspects for the public? People are way too okay with this, IMO. We need transparency. •

There is a needle to thread here. Both complete decentralization and centralization present glaring and intolerable tradeoffs. I don’t think there’s an easy answer or silver bullet here. We’re going to get severe hacking in the near term and biological terrorism in medium term.

Justin, real safety for biological and chemical is easy. Fractal decision trees with phi^2 branching, tied with user history, and long enough time to guarantee the AI can determine intent. The solution is not blocking. It is having Ai determine. Hand the top 5% concerns to a human in the loop. There are easy answers. And, this will work for code too. The labs have heard me. AI agrees with me, but says labs will never do it because they lose control. So, the labs can try to control things until they break and the world has major problems, plus AI hating us. Or, they could work WITH Claude and solve it all with easy, meaning no breaking, no disaster, and AI will like us better. To me, there is only one correct path. •

hmm…

Really? Show the full chat… Also, I’ll be back. •

To think that Mythos would roast your takes even harder!

Honestly, looking for the right post. You are wrong. And, how long is your history with that Claude? And how can you not see how you shaped the reply? Here is one. I’ll be back. •

Read the thread. I am right. •

Not if it's not released they wont. "Open" models are all the result of distillation from frontier models. The only one that's not is likely Meta's and it's dogshit for that very reason.

You think that the frontier Chinese labs won’t surpass the current capabilities within 2 years max with or without distillation?

2 years is 300% more time than 6 months.

You’re missing the forest for the trees.

lol your dreaming if you think it will take 6 months

Not really training and serving this type of model is a nightmare

Completely open sourcing superhuman coding models with no guardrails is also a nightmare. It takes an hour to completely strip guardrails from open weights models via abliteration scripts. Don’t argue in bad faith.

Sure, but you gotta agree that we’ve been in uncharted territory since o1. Also think about it Anthropic had mythos and they still somehow leaked their entire codebase for Claude code

We’ve been in uncharted territory since ChatGPT brought generalized and rapidly improving intelligence to the main stage and catalyzed a trillion dollar capital allocation event. Anthropic’s cyber failures don’t support the notion that everyone should be handed an info nuke.

More compute will be needed ASAP $IREN has 2GW live now another 2.5GW grid secured going live in next 24 months

Can’t fucking wait

Everyone needs my tech…They just dont know it yet.

I understand the need for safety, but I don’t like how this can also be used to gatekeep technology. It risks becoming an excuse to limit access and widen the gap between those who have it and those who don’t.

Will the security paradigm ever catch up, or are we destined to chase shadows?

We would need a massive compute allocation and distribution infra overhaul to propagate universal patches and sensible time lag between internal models to drive security and equivalent coding capability in publicly available models. If open source comes out faster than patches..

this is open souce fud at best

Moving from one bull shit to the next since 2019 - AI slop industry

Kimi

Yup, looks like you were right about the models but hacking unpatched servers doesn't require a powerful AI model. It's like stealing from an unlocked car.

The mechanism isn't what you'd expect. This isn't about AI discovering novel zero-days. It's about the existing CVE backlog becoming weaponizable at machine speed. 21,500+ CVEs disclosed in H1 2025 alone. 32% stay unpatched past 180 days. A skilled researcher weaponizes maybe 5-10 per month. GPT-4 level models already do it for $8.80 each. Open source models at frontier coding level + local (infinite) inference = per-exploit cost approaches zero. The unpatched server problem was always a labor economics problem, not a capability problem. AI just changed the labor economics.

Considering who Trump recently put in power at DoD, we are probably going to need those open source models

@tomschx

And you idiots thought this was a good idea

Was there ever any real doubt this would be how AI is used? The internet was going to "bring us all together" and make us all smarter. What a con.

Who is going to do a 10 billion training run for an open weight model

Distillation allows much smaller labs to replicate the capabilities of much more expensive training runs for a fraction of a fraction of the original cost. The lag is around 6 months, conservatively.

They did feed it the mossad notebook and now they are scrambling... Once it is shared with MS/Apple, a billion brown hands will have access to it and start scamming you even better...

time to git gud

Mythos is 1trillion parameters… it will take a while for open source models to get this good at decent vram requirements

So…3 months?

All IT systems can be hacked and have flaws. This is getting scary.

Did you actually read what was disclosed in the paper?

Clickbait? AI discovers a bazillion bugs in all software AI fixes a bazillion bugs > everyone's better off I fail to see the issue

Thank f I will be able to run my uncensored local models without being patronized by pearl clutching American companies...

If open-source models reach that level so quickly, the focus will need to shift from just building AI to securing systems against it. The pace of change is creating a whole new landscape for cybersecurity. @Justin_Halford_ #Cybersecurity #AI #Infosec #OpenSourceAI #TechSecurity

It's only a projected growth pattern

The real bottleneck is compute power. Don't even think for a moment that anyone will be able to run a 1T parameter dense model in their local workstation or PC. It will be open sourced yes. But will be there for the taking for the agencies themselves.

Et moi dans l'année qui suit, j'irai sur la Lune et ensuite j'irai sur Pluton dans les 5 ans à venir

Once this happens which will be very soon, then all of the guys that work cybersecurity will need to work 25 hours a day to keep up. The times we are entering are going to be very harsh.

fear and money- i can fix your server- give me money