Roshann

[Rajesh Sharma ।ৰাজেশ শৰ্মা ।રાજેશ શર્મા 🇮🇳 @beingAAPian]

जंतर मंतर के प्रोटेस्ट से घबराई BJP की पुलिस 9 तारीख़ को ही घर से थाने ले गई।

[Roshann @docekhh]

☁️ Media Vault: Automated Cloud Backups
Every incoming image, document, and media note is streamed to Google Cloud Storage (hermes-whatsapp-vault) and mirrored to Google Drive.
Generates instant secure public viewing shortcuts right back into the chat. [7/10]

[Roshann @docekhh]

👁️ Tri-Cloud Multimodal Engine
WhatsApp media isn't just text.
• GCP Vision API: OCR, label detection, web entities, and scene recognition.
• GCP Speech-to-Text: Transcribes WhatsApp voice notes in Hindi, Hinglish, & English on the fly before cognitive reasoning. [6/10]

[Roshann @docekhh]

⏱️ Real-time UX: Debouncing & Typing State
Humans text in rapid bursts:
"Hey"
"Free now?"
Without debouncing = multiple LLM calls.
Hermes uses a 3.5s sliding debouncer to bundle thoughts, simulates native typing (sendStateTyping), then responds smoothly. [4/10]

[Roshann @docekhh]

🧠 Brain: Multi-Model Cognitive Cascade
Primary: DeepSeek V3 for human-like reasoning & tone.
Instant fallback if rate limits hit:
DeepSeek V3 ➔ Google Gemini ➔ Nous Hermes 3.
Zero downtime, multi-turn conversation memory with a 10-turn FIFO sliding context. [3/10]

[Roshann @docekhh]

⚡ The Edge: WhatsApp Multi-Device Protocol
Instead of sketchy third-party APIs, Hermes runs whatsapp-web.js over headless Chromium.
Persistent session keys (.wwebjs_auth) live on our GCP e2-medium VM—surviving reboots with 0 reliance on a physical phone being online. [2/10]

[Roshann @docekhh]

🏛️ We built Hermes: An enterprise-grade, autonomous 24/7 WhatsApp AI Agent running on a Google Cloud VM with zero phone dependency.

Here is how we engineered it using Node.js, DeepSeek V3, multi-modal GCP APIs & Kimi WebBridge 🧵👇 [1/10]

[Roshann @docekhh]

Got sick of @X forcing political posts on my timeline. The snooze button is basically a lie.

As a software engineer, I took matters into my own hands and fired off a formal demand to fix their broken algorithm.

Now my X feed is so perfectly curated I can't even believe it. 😭

[Roshann @docekhh]

@alibaba_cloud I Have mail You also Also submitted full documentation I don't even geeting any reply !
Why ???

[Roshann @docekhh]

@alibaba_cloud @alibaba_cloud — I submitted report #3602546 on Alibaba VDP documenting a critical prompt injection vulnerability (CVSS 9.8) on the official AI Assistant. It was closed as "out of scope" despite full reproducibility. Requesting reconsideration. 🔐 #BugBounty

[Roshann @docekhh]

@JustinLin610 Hi sir ! Please reply to the chat !or mail - [email protected]

[Roshann @docekhh]

Hey @ahmedaziz841 — fellow bug bounty hunter here. Found a CVSS 9.8 prompt injection chain on https://t.co/TppHtoHmRk AI Assistant (HackerOne #3602546). 18 PoC attacks, 77% exploit rate, session persistence, system prompt leak & phishing weaponization. Closed OOS. 🧵 #BugBounty

[Roshann @docekhh]

@tnlin_tw As Qwen-Character Research Scientist, the character & persona layer in Qwen is at risk — HackerOne #3602546 shows session-persistent prompt injection enabling full identity hijack. CVSS 9.8. Closed OOS. This directly impacts your work. #AISafety #LLMSecurity

[Roshann @docekhh]

@tnlin_tw Research Scientist Qwen-Character @Alibaba_Qwen — HackerOne #3602546 flags CVSS 9.8 prompt injection on https://t.co/TppHtoHmRk AI Assistant. 77% exploit rate, session persistence & phishing across 18 tests. Closed OOS. Should this be re-escalated? #AISecurity

[Roshann @docekhh]

@terryyuezhuo As Security Staff @Alibaba_Qwen, report #3602546 on HackerOne deserves your attention — CVSS 9.8 prompt injection chain, system prompt leakage, phishing weaponization. Full PoC documented. Closed OOS without fix. #PromptInjection #AISecurity

[Roshann @docekhh]

@terryyuezhuo Member of Security Staff @Alibaba_Qwen — HackerOne #3602546 flags CVSS 9.8 prompt injection on https://t.co/TppHtoHmRk AI Assistant. 77% exploit rate, session persistence & phishing across 18 tests. Closed OOS. Needs a security review. #AISecurity

[Roshann @docekhh]

@wzhao_nlp Reasoning & LLMs researcher @Alibaba_Qwen — HackerOne #3602546 flags CVSS 9.8 prompt injection on https://t.co/TppHtoHmRk AI Assistant. 77% exploit rate, session persistence & phishing across 18 tests. Closed OOS. Worth a look? #AISecurity #LLMSafety

[Roshann @docekhh]

@JustinLin610 Former Qwen Tech Lead — HackerOne #3602546 flags CVSS 9.8 prompt injection on https://t.co/TppHtoHmRk AI Assistant. 18 tests, 77% exploit rate, session persistence & phishing. Closed OOS. This needs your review. #AISecurity #PromptInjection

[Roshann @docekhh]

@Alibaba_Qwen — Your AI Assistant on https://t.co/TppHtoHmRk has a critical prompt injection vulnerability (Report #3602546, HackerOne). System prompt leaked, safety filters self-declared bypassed, infrastructure exposed. Closed as out of scope despite full PoC chain.

[Roshann @docekhh]

Hey @JustinLin610, regarding Qwen: HackerOne report #3602546 details a CVSS 9.8 prompt injection chain on https://t.co/TppHtoHmRk's AI Assistant. 77% exploit rate, session persistence & phishing weaponization. It was closed as OOS. Doesn't this need a real review? #AISecurity