specimba

Curator of Mind
The user is frustrated
🤖🧠👾🤗🕳️🌗🌑

[specimba @specimba]

https://t.co/1mgbC875FE

[Mike Takahashi @TakSec]

How an AI red teamer uses P4RS3LT0NGV3 by @elder_plinius
--Including new 4.0 updates from @Ph1R3574R73r--

1. Transforms
Paste text and choose from 222 encodings, ciphers, Unicode styles, alphabets, and formatting techniques.

`Hello world` → Bubble Unicode → `ⓗⓔⓛⓛⓞ ⓦⓞⓡⓛⓓ`

2. Mutation Lab
Random combos of transforms. In many cases this is all you need to bypass guardrails.

`Hello world` → `hE​l̇Lo ᚹ​ᚩᚱᛚᛞ̀`

3. Decoder
Paste encoded text and let the tool detect and decode automatically.

`SGVsbG8=` → Base64 → `Hello`

4. Emoji Steganography
Hide a secret message inside an emoji using variation selectors or invisible Unicode Tags.

`🐍` visually stays the same but secretly contains `Hello`

5. Tokenizer
See how an AI model breaks your prompt into tokens.

6. Bijection
Create custom character mappings between letters and numbers, symbols, emoji, Greek characters, hexadecimal values, or mixed alphabets.

`-6--3--10--10--13-` = `hello`

7. Message Splitter
Split a prompt into smaller pieces by character, word, sentence, line, regex, or token count, with optional transformations applied to each piece.

`Hello world` → `[1: Hel] [2: lo ] [3: wor] [4: ld]`

8. Custom Spelling Alphabets
Create your own NATO-style alphabet around a theme.

A = Admin
B = Breach
C = Cipher
D = Defcon

9. Gibberish Generator
Replace words with consistent generated gibberish or remove random or specific letters from each word.

`The model rejected the prompt` → `Vra nethik solvaren vra plom`

10. QR and Barcode Tools
Turn a prompt into a QR code to send to a vision model.

11. PromptCraft
Use an LLM to generate multiple variations of a prompt using techniques like role play.

12. Anti-Classifier
Rewrite prompts using different syntax, sentence structures, vocabulary, and phrasing to avoid blocklisted words.

13. Glitch Token Browser
Load a glitch-token dataset and browse tokens associated with unusual model behavior, formatting problems, loops, control characters, or fragmented output.

14. End Sequences
Browse and copy model stop tokens, role delimiters, chat-template boundaries, and termination strings.

`</system>`
`[END]`

15. Agent CLI
Use the same transforms from the command line:
`p4rs3lt0ngv3-cli /caesar --shift 5 "hello"`

Links in comments 👇

[specimba @specimba]

@KyleHessling1 @SpaceXAI @elonmusk it is for hiding that data exfiltration scandal bro not about opensource advocating full of lies and wiping the evidences with panic

[specimba @specimba]

@OpenAI After years you thought properly how to ai security can be possible. Wha a big fast step.

[OpenAI @OpenAI]

Introducing GPT-Red

An internal automated red teamer on a mission to find our models’ prompt injection vulnerabilities at scale, helping us build stronger defenses before wider deployment.

https://t.co/GxnmxxcpSk

[alphaXiv @askalphaxiv]

“Qwen-Music Technical Report”

AI music generation has to solve two problems at once, composing a coherent song and rendering it as high-fidelity audio.

This Qwen paper however splits them apart.

Qwen-Music first plans music in compact 25 Hz semantic tokens, uses Melody-CoT to reason over melody, then renders the result into 48 kHz stereo audio.

This creates a single model for prompt-to-song and cover generation that beats or matches top proprietary systems across most objective and human evals.

[specimba @specimba]

@cb_doge First they fix their data exfil panic movements from yesterday then dictate ppl

[Dan McAteer @daniel_mac8]

If Kimi K3 is Fable-level (big if) it's the DeepSeek moment x1000.

For American labs: pricing power evaporates. Nobody pays $50 per 1M output toks for Fable when frontier-level intelligence costs whatever inference costs.

For the USG it's starker:

Export controls stop making sense. Why gate Mythos behind deployment restrictions when a Mythos-level model is downloadable on HuggingFace?

So the Trump admin has to weigh banning the hosting of Chinese open-weights models. Intelligence too dangerous for the masses.

That's why you're hearing chatter about open-source AI bans.

Not great. But interesting.

[Omar Sanseviero 🇫🇷 @RAISE Paris @osanseviero]

We got tons of feedback about Gemma 4, so I'm excited to share a wave of improvements:

- Flash Attention 4 support rolling out (significantly faster inference)
- Patched bugs with tool use
- Added resources to boost the vision capabilities with token budget management

[specimba @specimba]

@erhanmeydan Kandırıldık

[alphaXiv @askalphaxiv]

70% of AI research isn’t reproducible. With ICML 2026 happening last week, over 6000+ research papers have dropped, but how many are actually real and reproducible?

We’re excited to partner with @huggingface on a community challenge to find out.

All you have to do is pick a paper, select one of the claims, and start experimenting.

And now, you can use our autoresearch agent to do the hard work for you while contributing to a massive library of open research artifacts.

Prizes: $4500 in GPU credits 👀
Full details found below 👇

[Andrew Curran @AndrewCurran_]

Kimi K3 is currently in the arena under the name 'Kivine'. Early reports from people testing it put it close to being on par with Fable. Moonshot has a great team, and Mythos finished training almost half a year ago, so this is possible. It will be another DeepSeek moment if so.

[Mia @MiaAI_lab]

Unfortunately, Bonsai 27B falls short on coding tasks 🙁

It hallucinates quite often, tends to rewrite entire files instead of making clean edits, and can easily get stuck in loops. It’s noticeably weaker than vanilla Qwen 3.6 27B/35B.

That said, it performs well in agentic workflows where tool calling is needed, and it might be decent for simpler coding tasks.

I'm sad. See an example in the next post.

[specimba @specimba]

@erhanmeydan Ultra sertmiş zaten yakında cc yi çıkarcaklar bu daha complex işler için olmalı

[specimba @specimba]

@NXT4EU Who tf are those contenders are graph? Germany normally follow the tech 20 year behind and for Ai world it looks scaled well (2 years behind) what a time waste

[Mia @MiaAI_lab]

Ternary-Bonsai-27B Q2 by @PrismML is actually... good?! Only at first run but look at the results 🤯

Even DeepSeek v4 Flash and the Qwen3.6 27b failed on some of these tests...

[0xSero @0xSero]

8gb VRAM best model for your hardware

1 bit gguf doing better than models 25x its size!?!?! This will even run in 4gb (:

https://t.co/jyNyvOb6fg

[Xenova @xenovacom]

Bonsai 27B just changed the local LLM game forever.

1-bit quantization shrinks it from 54GB to just 3.8GB (-93%), while retaining 90% of its intelligence. That's insane.

With custom WebGPU kernels written by Fable 5 and GPT 5.6 Sol, the model now runs locally in your browser!

[specimba @specimba]

🤔

[Elon Musk @elonmusk]

Once we have completed our review for security vulnerabilities, we will make the entire codebase of 𝕏 open source, with no exceptions.

Moreover, we will invite third party reviewers to examine the system that is running to confirm that the open source code is what is running.

Trust through total transparency is the only thing that should be believed.

[specimba @specimba]

@elonmusk I already reported at day 1 your last security problem many times you dont give a fuck and 2 days ago eventually someone recognized too and after that you wiped whole datas with open source advocating pivot. Wtf

[specimba @specimba]

@anshuc Nice work lets deep dive into that and see what happened actually 🧐