AI agent truy cập Gmail có thể bị tấn công Prompt Injection, dẫn đến rò rỉ dữ liệu nhạy cảm. Giải pháp: dùng Model Armor của Google Cloud để bảo vệ, lọc nội dung độc hại và bảo vệ riêng tư người dùng. Kết hợp Model Context Protocol (MCP) giúp tách biệt an toàn giữa LLM và dữ liệu. #AI #Security #Gmail #ModelArmor #PromptInjection #AIAnToan #BaoMatDuLieu #GoogleCloud #LLM #EthicalAI

https://dev.to/googleworkspace/securing-gmail-ai-agents-against-prompt-injection-with-model-armor-4fo

"From Prompt Injections to Protocol Exploits: Threats in LLM-Powered AI Agents Workflows"

In this survey, we introduce the first unified, end-to-end threat model for LLM-agent ecosystems, spanning host-to-tool and agent-to-agent communications, formalize adversary capabilities and attacker objectives, and catalog over thirty attack techniques.

https://www.alphaxiv.org/abs/2506.23260v1

#research #AI #LLM #promptInjection #MCP #threatModel #taxonomy

Poetry as a cybersecurity threat? Yes, really.

A new paper shows “poetic jailbreaks” can bypass AI safety filters far more often than prose:

• 62% success with handwritten poems
• 43% with generated verse
• 8% with equivalent prose

Style, rhyme & metaphor confuse safety heuristics – the model “sees” a cute rhyme, not a threat.

Paper: https://arxiv.org/pdf/2511.15304

#AI #Cybersecurity #LLMs #PromptInjection #MachineLearning

CrowdStrike rolls out Falcon AI Detection and Response as AI prompts become the new attack surface

https://fed.brid.gy/r/https://nerds.xyz/2025/12/crowdstrike-falcon-ai-detection-response/

AI-powered browsers are opening up a whole new attack surface: indirect prompt injection. Malicious instructions can hide in web pages, PDFs, emails—even after the “#” in a URL. Your browser’s AI assistant doesn't know better. #CyberSecurity #AIBrowsers #PromptInjection

El lado del mal - Qué hacer para desplegar Inteligencia Artificial con seguridad en una empresa https://www.elladodelmal.com/2025/12/que-hacer-para-desplegar-inteligencia.html #IA #AI #Cloudflare #WAF #AISecurity #DLP #Guardrrails #PromptInjection #ThreatIntelligence #LLM #Jailbreak

El lado del mal - Ciberseguridad e Inteligencia Artificial: Mi última charla de 2025 la tienes en Youtube https://www.elladodelmal.com/2025/12/ciberseguridad-e-inteligencia.html #IA #AI #Ciberseguridad #conferencias #charlas #Eventos #hacking #Privacidad #LLMs #PromptInjection #Jailbreak

Who didn't see this coming from a mile away? #AI #AntiGravity #Google #PromptInjection
https://www.youtube.com/watch?v=S4oO27tXVyE

"Quicklines and Silly News - Weekly News Roundup 12/11/25" 👀👏

https://friendica.world/display/84b6ef2b-1069-3c20-c3cc-585561902992

El lado del mal - GeminiJack: Indirect Prompt Injection en Google Gemini Enterprise https://www.elladodelmal.com/2025/12/geminijack-indirect-prompt-injection-en.html #Gemini #PromptInjection #Google #GeminiJack

Nothing new but a good article on #promptinjection by #ncsc. While sql injections despite being common are mitigatable, prompt injections are not and are more dangerous. #itrisk #infosec #llm #genai

https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection

The Normalization of Deviance in AI

#ai, #ai-ethics, #generative-ai, #johann-rehberger, #llms, #prompt-injection, #security

PromptPwnd – nie zgadniecie co się okazało po integracji LLM z GitHub Actions

Gdybyśmy otrzymywali dolara (no dobra – starczy 1 Polski Nowy Złoty) za każde nieudane połączenie AI z czymkolwiek, to właśnie pisalibyśmy ten tekst z pokładu jachtu. Tym razem “nowa” odsłona ataku została nazwana PromptPwnd (standardowo brakuje tylko logo i dżingla), a dotyka ona potoków CI/CD (ang. Continuous Integration/Continuous Development pipelines)....

#Aktualności #Ai #Cicd #Github #Llm #PromptInjection

https://sekurak.pl/promptpwnd-nie-zgadniecie-co-sie-okazalo-po-integracji-llm-z-github-actions/

https://winbuzzer.com/2025/12/09/google-unveils-security-architecture-to-shield-chrome-ai-agents-from-prompt-injection-xcxwbn/

Google Unveils Security Architecture to Shield Chrome AI Agents from ‘Prompt Injection’

#AI #Cybersecurity #GoogleChrome #AIBrowsers #AIAgents #AgenticAI #PromptInjection #GoogleGemini #BrowserSecurity #InfoSec #OpenAI #Perplexity

New AI-native threat: GeminiJack allowed zero-click exfiltration of Gmail, Docs & Calendar data in Google Gemini Enterprise + Vertex AI Search via indirect prompt injection in the RAG pipeline.

Analysis:
https://www.technadu.com/new-ai-native-threat-vulnerability-in-google-gemini-enterprise-and-vertex-ai-search-allowed-stealing-gmail-docs-and-calendar-data/615399/

#GeminiJack #AINativeThreats #GoogleGemini #VertexAI #PromptInjection #RAGSecurity #CyberSecurity

UK intelligence warns #AI #PromptInjection attacks might never go away

https://therecord.media/prompt-injection-attacks-uk-intelligence-warning

In case you need another reason to hate #LLMs , the shittiest of #ArtificialIntelligence #tech.

#security #cybersecurity #InfoSec #technology #LLM #LargeLanguageModels

Google details layered security for Chrome’s agentic AI, featuring a Gemini-powered "Critic."

Google is safeguarding Chrome's upcoming autonomous AI features with a "User Alignment Critic" model that vets every action for user safety, Origin Sets to prevent data leaks, and mandatory user confirmation for sensitive tasks like payments and banking access. These measures are designed to combat indirect prompt injection attacks.

#chromeai #agenticai #gemini #browsersecurity #promptinjection
9m

Read the article 👉 https://www.giskard.ai/knowledge/agentic-tool-extraction-multi-turn-attack-that-exposes-the-agents-internal-functions

#AIAgents #PromptInjection #LLMSecurity

"security teams and those owning the risk need to be aware that #promptinjection attacks will remain a residual risk, and cannot be fully mitigated with a product or appliance etc. It needs to be risk managed through careful design, build, and operation."

https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection