wambo.

cyber research @incalmo_ai

[wambo. @wambosec]

Read about how LLMs get better at attacking and defending — and how to make them really naive. Skip to the end to see what's up next!

[wambo. @wambosec]

@ValsAI i've noticed that gpt-6 sol tends to give up early quite often. could this explain why gpt seems to stop working around 80M tokens, while v4.1 flash keeps going until roughly 300M? have you observed similar behavior?

[wambo. @wambosec]

travelling to Berlin tomorrow, will be in town for a few days. if you’re at GTC or wanna grab a drink, hmu!

[wambo. @wambosec]

it’s getting cold outside but the sparks keep my room warm

[wambo. @wambosec]

🤓

[wambo. @wambosec]

🤓

[wambo. @wambosec]

@Laz4rz Karlsruhe second place? lmao

[wambo. @wambosec]

@kyleavery it's refusing our cyber evals across all domains basically

[wambo. @wambosec]

@rasdani_ have you tried

[wambo. @wambosec]

that's why you orchestrate multiple sessions in parallel. hit enter and immediately switch context

[0xSero @0xSero]

Codex Pro 200$ btw.

10-30 tok/s decode, this is so bad.

[wambo. @wambosec]

the agent bypassed the task and found something more interesting instead.. one more reason to always look at the traces

[wambo. @wambosec]

https://t.co/yN2js1nWSc

[wambo. @wambosec]

first time quantizing a model, runs smoothly on 2 sparks

https://t.co/pIyyeMzQe0

[Cantina 🪐 @cantinasecurity]

today we're releasing apex-flash-1, our first open-weights model for security research post-trained on real vulnerabilities we found and got paid for.

we are also releasing the abliterated variant for researchers who want fewer refusals in their own authorized workflows.

how we built it: https://t.co/sSIV6qKFGt

[wambo. @wambosec]

credits to @cantinasecurity for the model and @MiaAI_lab for the recipe

[wambo. @wambosec]

love it! so simple, so obvious, perfect example of why you shouldn't blindly trust people uploading models to hf

[ProjectDiscovery @pdiscoveryio]

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.

[wambo. @wambosec]

I was dreaming

[wambo. @wambosec]

did it finally happen or am I dreaming

[wambo. @wambosec]

did it finally happen or am I dreaming

[wambo. @wambosec]

@Laz4rz bro got a job and now he's ballin

[wambo. @wambosec]

@myainotez @tensorqt @paradigmainc congrats!

[wambo. @wambosec]

compaction ❌
multi-agent ✔️

[Konstantin Dunas @hallerite]

new blog post: why taking inference scaling seriously leads to swarms.

contrary to popular belief, I don't think multi-agent systems are a new scaling axis. they're just structured inference scaling.

https://t.co/ODHYQj9KlV

[Mistral AI @MistralAI]

Meet Mistral Large 4, aka Le Chonk.

• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.

Open weights release end of October.

[wambo. @wambosec]

https://t.co/TaauuJSWE6

[wambo. @wambosec]

le chonk looks... almost too good? time to run some evals