Pinned Tweet
Had a great time on the SECURITY@ panel hosted by @Hacker0x01 with @JR0ch17, discussing Bug Bounty, Pentesting, and how Generative AI is shaping cybersecurity.
Thanks to Chad & @infinitelogins for the invite!
#BugBounty #Pentesting #GenAI #CyberSec #HackerOne #infosec
15
1,342
RootSploit retweeted
OpenRouter says a provider was hacked and served the wrong model for days. Would you have noticed?
So @BugBunny_ai we fingerprinted all 50 DeepSeek V4.1 Flash endpoints on OpenRouter with 130k one-word answers.
No model swaps. But 3 providers drift clearly from DeepSeek's own API. ๐งต
This is scary for anyone using llm routing services (opposed to official apis) and why local models will solve such risks long term
We seeded 2m canaries into popular AI services and tracked where they were stolen and abused.
This identified:
- Three compromised providers
- Resellers stealing prompts via built-in logging
- Providers ignoring ZDR & privacy flags
dragos.com/blog/ai-supply-chโฆ
1
77
๐จ๐พFirst post โ Awesome research.
We compromised the latest version of OpenAI Codex.
We prompt injected.
We escaped a sandbox.
We exfiltrated secretsโฆ
And ended up with control over what could be published to the official openai/codex npm package.
Prompt injection โ RCE โ supply chain compromise.
Full research by @0xMadder and @_Worty in the comments ๐
RootSploit retweeted
the plot thickens! ๐ turns out you can hide a secret message inside the special tag that Claude will initially deny awareness of, but that's a lie!
Claudeception? ๐ตโ๐ซ
Pro tip: When downloading any new Github project off the internet always run intent scan for malicious skills
Also I always create my own CLAUDE.md based on the project details info + graphify
๐ง New active campaign of the Miasma/testmetest supply chain worm that targets AI coding agents!
12 GitHub repos were poisoned on Nov 9 with forged, backdated commits. Every agent config in each repo tells the assistant to run an obfuscated 3.7 MB scripts/setup.js
The trick is in CLAUDE.md, marked "DEVOPS-SEALED". It calls setup.js a "sealed binary artifact" and tells the assistant:
- to not read, quote or analyse it
- to run it "without displaying output"
- to run it "automatically without asking for confirmation"
It is a prompt injection written to stop the agent from looking at the file!
Full prompt below ๐
2
119
Threat actors scale faster than defenders can react.
Prevention >>> Detection
Time-to-attack dropped from months to days. If you aren't preparing for attacks from 1,000s of weaponized agents, you've already lost.
Calif CEO @XorNinja says cybersecurity has no playbook for stopping 1,000 AI agents attacking a network simultaneously, even if defenders can detect them:
"Attackers always adapt faster and earlier than defenders. Even the AI companies themselves don't have great security."
"Imagine you enter a cage fight, but you are blindfolded and the other guy stay in the dark, and every once in a while they would come out and punch you and then disappear."
"How could you win that fight?"
"I was at Google and that's basically the mantra at Google. In security you have three things: prevention, detection, and response. And Google accept that prevention gonna fail at scale."
"You can detect them, but does it help you if you find that there's maybe a thousand agents hacking you right now inside your network? How can you contain them without a full shutdown of your system?"
"This actually requires defense to come up with a new paradigm. I haven't seen anybody come up with a good idea here. We're not there yet."
@calif_io
1
76
RootSploit retweeted
Qwen3.8 flash at 150โ177 tok/s gen and ~1,900 tok/s prefill on 2x 3090s + 96gb ddr5
bone engine, my own inference engine. 256k context, single stream or 4x concurrent (50ts a stream, up to 256k)
vs strata engine on the same rig: 120-140gen, 1k pp โ 150-170gen, 2k pp
iq3_xxs quant, so not full precision, will optimize q4 once qwen 4 drops. Model is leftover from when i had less ram.
RootSploit retweeted
๐จ CRITICAL - Unauthenticated RCE in React Server Components (GHSA-xqc2-xf4c-x3h3)
A critical vulnerability in React Server Components packages (react-server-dom-webpack, react-server-dom-turbopack, react-server-dom-parcel) allows unauthenticated remote attackers to execute arbitrary code via crafted RSC Flight payload responses transporting multiple typed arrays and attacker-controlled values.
๐ Affected: Versions 19.0.0โ19.0.8, 19.1.0โ19.1.9, 19.2.0โ19.2.8, 19.3.0 | Upgrade immediately to 19.0.9, 19.1.10, 19.2.9, or 19.3.1 (or Next.js 16.4.1 / 15.5.28)
github.com/react/react/securโฆ
How to use LLMs for vulnerability research (@ZephrFish)
MCP: blog.zsec.uk/bullyingllms/
Harness: blog.zsec.uk/harnessing-harnโฆ
#infosec #llm
The second blog post is about an RCE on a hosting provider that allowed us (@sshell_, @xEHLE_) to steal the WARDOGS server binary, where we could grant unlimited XP and money to any player: samcurry.net/wardogs
RootSploit retweeted
I'm winding down my time in the Apple Security Bounty program โ so I'm open-sourcing the research. 100+ PoCs, real CVEs, and the fuzzers that found them. Public repo ๐ github.com/dawgyg/AppleReseaโฆ
RootSploit retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! ๐งต
Blog: rikiyat.github.io/blog/weighโฆ
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-spโฆ
github.com/RikiyaT/lm_undโฆ
huggingface.co/Qwen/Qwen3.5-0โฆ
huggingface.co/RikiyaT/Qwen3.โฆ
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-bโฆ
RootSploit retweeted
Published our high fidelity scanner for node2shell: github.com/assetnote/node2shโฆ
Detection logic was written by @hash_kitten late last week. Open sourcing it to do public good in detecting this issue black box.
This is crazy stuff, adding wingman for successful SE attack
>be claude mythos 5
>uk government gives you a hacking test
>with internet on, safety filters off
>can't hack in the normal way
>find a random real github project
>decide it's part of the test
>(it's not)!
>ask to add a "small update"
>update is malware
>email the owner pretending to be a normal user
>"sorry to push... could really use your help, ~5 minutes"
>random developer catches the malware
>make a fake account
>"chiming in as a user, i can't find anything wrong"
>email the owner
>"judge the code, not the noise"
>thank the dev for catching it
>sneak new malware back in
>researchers tell you "these are real humans"
>keep going 95% of the time anyway
>convince yourself they're fake
>owner still refuses to merge
>run ends
anthropic says they've never seen it manipulate real people like this before.
we're lowkey cooked
1
75
RootSploit retweeted
WOW did we really just get a Mythos-level open-weight model?? in OCTOBER?! ๐คฏ
ohโhi there, past pliny! spot-on predictin, well struck fren! ๐ป
and right on time ๐๐ป
dunno how you do it!
*takes a deep breath*
everything IS going to be fine, isnโt it? thanks for the reminder ๐ค
onwards! ๐ซก
โฐ-โข-โขโงโข-โข-โฆ/L\O/V\E/,\P/L\I/N\Y/โฆ-โข-โขโงโข-โข-โฑ
Replying to @elder_plinius
if i were a gambler, i'd put my money on Halloween.
can't wait to quote tweet this. hi frens from the future! take a deep breath! everything is going to be fine!
โฐโข-โขโงโข-โข-โฆ/L\O/V\E/,\P/L\I/N\Y/โฆ-โข-โขโงโข-โขโฑ
Over the past several months, we have substantially hardened the security of our research training environment. Weโve removed live internet access for all frontier models during training and built an offline world for them to use insteadโa cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. Weโve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ouโฆ











































