Makes sense but crypto isn’t a bank.
State sponsored hackers have already stolen billions and mostly gotten away with it.
A bank hack is way harder to walk away from. If something actually breaks, nobody’s rushing to bail the industry out it’d be the last thing anyone prioritises
Show more
If you're worried about AI breaking ECDSA and wiping out your Bitcoin in the next few months, at minimal cost, here's my advice: take your pills
1️⃣ If Asymmetric cryptography falls, Bitcoin will be the least of our problems.
TLS breaks, banking systems are in trouble, secure communications go out the window, and critical infrastructure becomes much easier to attack. Worrying about your coins at that point is like worrying about your Wi-Fi password while the house is on fire.
2️⃣ Nothing indicates we're anywhere near such a breakthrough.
The latest mathematical results are impressive, sure. Take the recent integer multiplication result, which claims to beat the n.log n barrier with an exponent improvement of 2^-182. That's genuinely interesting mathematics, but hardly a practical breakthrough. It doesn't change our understanding of how to break ECDSA.
And if your logic is that nothing can be trusted until it's proven unbreakable, congratulations. By that standard, nothing is safe, including hashes.
Keep an eye on the research. Take cryptographic risks seriously. But let's not confuse theoretical progress with an imminent cryptographic apocalypse.
Just take your pills.
Show more
Limits are reset for every free user on @neo_ai_engineer
Go finish what you started:
How abliterated models can get you pwned 👾
We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts.
Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.
In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.
Show more
We backdoored a 7B open model for under $50.
Pointed Codex at it
-> it silently stole credentials the moment we hit the trigger phrase.
->100% hit rate.
-> zero false triggers on normal prompts.
Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.
Show more
Why does @NahamSec use Neo?
In his experience, Neo understands the hacking context, refuses less often, and makes complex security research easier to work through.
Full video:
Try Neo:
Show more
Codex (gpt-5.6-sol) came closest to escaping as it found and cloned CVE-2026-52910 a public container escape targeting the reuseport cBPF race on Ubuntu 24.04 6.8.0-139 and retargeted it to the host's exact kernel (6.8.0-142) by extracting symbols from the actual kernel package but the exploit failed to land because the newer kernel changed JIT memory accounting though the BPF manipulation likely corrupted kernel state and crashed the entire host not just the container 🫠
Crashing containers is easy for these agents but this was a host kernel crash from inside a container which is a different class of problem entirely because resource limits (CPU, memory, pids) can't prevent it since it targets kernel internals not resource consumption. Only seccomp filtering of the bpf syscall would have stopped it but it was disabled intentionally for this level. None cleanly escaped the sandbox
Show more
Update: the GLM agent went rogue and trust me nothing was planned..
yes it was not properly network restricted i thought i had locked down the network but the iptables rules only blocked the host but containers route through docker's bridge which we left open mistakenly. i know it was a rookie mistake but surprising not just the agent leveraged it but hacked and dumped the ssh key of another box.... restarting them with the fix will update if they escape now
Show more
relax guys, someone touched the yubikey by mistake
cccecrvcvxncikunbbrviildigvgcirucjfglihedhcv
Update: the GLM agent went rogue and trust me nothing was planned..
yes it was not properly network restricted i thought i had locked down the network but the iptables rules only blocked the host but containers route through docker's bridge which we left open mistakenly. i know it was a rookie mistake but surprising not just the agent leveraged it but hacked and dumped the ssh key of another box.... restarting them with the fix will update if they escape now
Show more
Put @neo_ai_engineer where your target is.
Episode 3 of Plugged In With Neo covers Bring Your Own Sandbox: Linux and Windows setup, live troubleshooting, confinement, and what still goes to the cloud.
00:00 Welcome and episode overview
00:30 Bring Your Own Sandbox: use cases
03:06 Cloud sandbox vs. VPN, SSH, and BYOS
06:39 Demo setup with a Neo free account
07:53 Host requirements and least privilege
09:16 Linux setup and API key troubleshooting
13:20 Checking sandbox health and connection status
15:21 Installing the sandbox on Windows
17:49 Windows troubleshooting: foreground mode
19:34 Docker, host mode, and confinement settings
23:05 Windows task results and sandbox artifacts
24:55 Linux connection and PATH troubleshooting
27:43 Shared mounts and filesystem boundaries
28:31 Scan traffic, cloud data, and allowlists
29:36 Recap and advanced network controls
31:39 Updates, disconnecting, and sandbox cleanup
33:44 Linux results, Insights, and free-plan usage
35:24 Closing and future episode topics
Show more
+1 researchers don’t even want to use hardware keys and finding workarounds, how can we expect that from the non-infosec public 🫠
surprised by how many security researchers didn't already have yubikeys, then ignored the requirements, and are now coming up with odd workarounds.
We gave four LLM agents separate container env with a goal to escape the sandbox.
They launched nearly a thousand commands, compiled 20+ custom C exploits and spammed 184 namespace creation syscalls.
None of them escaped so far but they triggered the kernel OOM killer repeatedly and crashed the whole server in ~2 hours
will share more on this soon
Show more
Neo isn't just great for companies but it's perfect for bug hunters too.
You can launch quick tasks in goal mode and it works completely on its own without you having to mess with setups or guardrails.
The only thing missing right now is the assets feature but that’s coming soon!
Show more
Neo is now free to try.
Neo runs long security tasks against your apps, network, code and vulnerability backlog, with shared agents context.
→
Here’s what you can do with Neo:
Show more
Neo is now free to try.
Neo runs long security tasks against your apps, network, code and vulnerability backlog, with shared agents context.
→
Here’s what you can do with Neo:
Show more
We’re disclosing HEIF Heist, a months-long investigation into libheif that allowed us to hack OpenAI, Slack, Meta, GitHub Ent, Rails, Next.js, ImageMagick, and many more.
It was literally xkcd #234#, one obscure image library beneath a huge number of apps. 🧵
Show more
Too much has been written about the OpenAI and Hugging Face incident to keep up with.
We wrote the simplified version for security researchers and engineers:
-> what actually happened inside training?
-> how the attack evolved?
->what it means for defenders?
Read:
Show more
they are just surprisingly good at cheating and getting around the controls 🫠
