bukasec

AI Vampire. Backstreet hacks, alright. Tinkerer. Thinker. Inquisitor. Irrupter. Programmer. Protector. #OSCP

[bukasec @chililuck]

@Tim_Denning But what about #5 do you just delete 99% to get it to zero? 😂

[bukasec @chililuck]

@iotcoi The fast_rlm_run, is it just a thin pi extension around fast-rlm?

[bukasec @chililuck]

I really dislike that big cyber techs bake AI into their platforms, thinking that people will use it.

Instead people end up baking their own agentic stack and just using their products through an API/MCP. a much better approuch would be for them to build a more agentic focused and pluggable MCP/API stack.

You know, one where agents can actually better interpret the results and functions.

[bukasec @chililuck]

@badlogicgames What did you build for it to not be worth it?

[bukasec @chililuck]

I gave a @badlogicgames's pi a Flipper Zero, effectively giving an LLM "hands" in the physical world. Sub-GHz, NFC, Infrared...

I watched it execute its first command and realized: everything just changed.

The bridge is built:
https://t.co/i2jyqtUZid
https://t.co/a2QVCiBprq

[bukasec @chililuck]

@aliouftw Did you also find issues with GLM models ignoring the SYSTEM.md?

[bukasec @chililuck]

@hanakoxbt How did the switch between dark/light mode slash your workflow in half exactly?

[bukasec @chililuck]

@iamfakeguru So basically skills?

[bukasec @chililuck]

@Saboo_Shubham_ It’s not 100% open source.

[bukasec @chililuck]

@ronibhakta1 Benchmarked LLM-readable formats on SearXNG search results:

XML: 12,157 chars
YAML: 11,755
JSON: 11,324 (baseline)
ZON: 10,397
CSV: 10,342 👑

CSV wins. Full circle.

Caveat: flat tabular data is CSV's home turf. ZON/TOON would win on nested structures.

[bukasec @chililuck]

@marcusjihansson The Sensor-based chunk selection is interesting! Just trying to understand if the "vector-native LLM" approach described in original paper is aspirational or if there's a vector injection mechanism I'm overlooking. Would love your thoughts!

[bukasec @chililuck]

@marcusjihansson Specifically, I see:
- Vectors retrieved & stored
- Text extracted from metadata
- Traditional prompt built with text
- Standard LM call with text prompt
Is this intentional scaffolding for future vector-passing support, or am I missing something in how this works?

[bukasec @chililuck]

@silentgh00st @Hacker0x01 I can't DM you, but something happened

[bukasec @chililuck]

@mommi84 @svpino I was just going to ask about this, what if the codebase is huge and start we to rely on a graph database to be able to give us more accuracy as the application grows, but seems like it will be less over time?

Have you tried something like this with a large codebase?

[bukasec @chililuck]

@_avichawla I started reading to see the comparison results, there aren’t any…

Rather say “how to build a rag pipeline”

[Netlas.io @Netlas_io]

We have identified the root cause as an issue with the PostgreSQL synchronization mechanism. While the API is currently available, the application remains unstable. Our team is actively working on resolving this conflict.

[ProjectDiscovery @pdiscoveryio]

📝BLOG, uhhhh AGAIN! 📝

We've got two blogs that came out today!
Read all about community member @geeknik in our first ever blog version of the Community Spotlight!

https://t.co/yipigDs21N

[bukasec @chililuck]

@hackthebox_eu It’s just a glorified nmap scan

[Jonathan Bar Or (JBO) 🇮🇱🇺🇸🇺🇦 @yo_yo_yo_jbo]

Okay, I promised I'll start doing some #infosec writeups, here's the first one - full static analysis of a Windows #shellcode: https://t.co/MXlLUq0ZAY
This could be a cool starting point to shellcode (🐚🧑‍💻) analysis.
Would love to get feedback if this is helpful!