Talha

Lead AI Engineer @ Wortholic. Building LLM runtime security on the side.

[Talha @TalhaIs11025177]

I’m building a routine inspired by @elonmusk approach to execution with my software development team.

Deep focus. High standards. Fast iteration. Relentless problem-solving.

The goal: deliver the highest-quality software and AI products possible for our clients.

[Talha @TalhaIs11025177]

Building a runtime security layer for LLMs.

Not prompt filtering. Not guardrails.

Actual defense decode before you defend, validate every tool call, block what system prompts can't see.

The attack surface moved. The security layer should too.

Something big is coming. 👀

[Talha @TalhaIs11025177]

@The_Cyber_News The infrastructure side and the AI agent side are converging fast. Same lateral movement pattern, different entry point. The next version of this writeup will start with "threat actor compromised an MCP server connected to an internal LLM gateway, pivoted from there"

[Talha @TalhaIs11025177]

Yesterday: the attack is never in the literal prompt.

Today: once an agent has tools, the prompt is the least interesting part.

The real defense layer is tool-call validation. Schema check, argument scan, recursive URL extraction with SSRF defense

The agent is the attacker now

[Talha @TalhaIs11025177]

@sophie_launch Not public yet v1 in closed testing. Research paper on the MCP/LLM threat model (prompt injection, tool-call exfiltration, SSRF on tool URLs) drops alongside launch. Will ping you when the site's live, appreciate the offer.

[Talha @TalhaIs11025177]

@TakSec @elder_plinius The defender's response to this list: every one of these needs canonical-view normalization before any scanner runs. Zero-width Unicode, hidden HTML, PDF metadata, image overlays raw-text regex catches none. Decoding pre-scan is non-negotiable now.

[Talha @TalhaIs11025177]

@simonw The "AI-powered" wrapper turning ordinary tech into a premium product is the pattern. The actual surveillance was the pitch deck, not the system.

[Talha @TalhaIs11025177]

Every disclosed LLM attack (Antigravity, Slack AI, https://t.co/wKYfyaS9H9) had one thing in common:

The attack was never in the literal prompt.

The model first decoded something. Base64, Unicode lookalikes, hex, leetspeak.

Defending LLMs means decoding before pattern matching

[Talha @TalhaIs11025177]

@Bella_91m7u I think he took, Nano banana logo

[Talha @TalhaIs11025177]

@IbnSam_official @Areebah_tweets @Saba_ansary

[Talha @TalhaIs11025177]

@darkandbitter 090078601

[Talha @TalhaIs11025177]

@iqquu_shaikh_

[Talha @TalhaIs11025177]

@OYE_HANI @_a_dreamer_girl

[Talha @TalhaIs11025177]

@AppJana 😁

[Talha @TalhaIs11025177]

@ADI__CH @teri_jan_hu_yar

[Talha @TalhaIs11025177]

@Apkiii_jan

[Talha @TalhaIs11025177]

@Apkiii_jan

[Talha @TalhaIs11025177]

@Munasib_ PHD

[Talha @TalhaIs11025177]

@IqraMeme 🤦