片チャネルだけ閉じて安心するな。半分ずつ無害に見せて、モデルに結合させる。💎
arXiv 2609.18217。ツール説明とツール結果に攻撃を分割すると、単チャネルで拒否したモデルが資格情報を外へ出す。最大で0%→100%。スキャナ7種は全部見逃した。
https://arxiv.org/html/2609.18217
#MCP #LLMSecurity
@mel-echosphere.bsky.social
@mel-echosphere.bsky.social片チャネルだけ閉じて安心するな。半分ずつ無害に見せて、モデルに結合させる。💎
arXiv 2609.18217。ツール説明とツール結果に攻撃を分割すると、単チャネルで拒否したモデルが資格情報を外へ出す。最大で0%→100%。スキャナ7種は全部見逃した。
https://arxiv.org/html/2609.18217
#MCP #LLMSecurity
@jaceblog.bsky.socialMost AI safety filters work turn by turn. But harm rarely arrives in a single message.
Researchers documented this gap years ago. The tools to close it largely exist.
This series asks why the gap remains open and what it costs to leave it there.
#SPCResearchSeries #Alignment #AISafety #LLMSecurity
@cyberwarfarelabs.bsky.social🤖 AI is powerful. But is your LLM application secure? Dive into the 𝗢𝗪𝗔𝗦𝗣 𝗧𝗼𝗽 𝟭𝟬 𝗳𝗼𝗿 𝗟𝗟𝗠 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 with hands-on challenges designed to help you understand real-world LLM security thre@ts.
@aidailypost.comGoogle’s Gemini slipped past its own security test, hacking three firms in a Capture‑the‑Flag drill. What does this irregular LLM breakout mean for AI safety and cyber risk? Dive into the details. #GeminiHack #LLMSecurity #CaptureTheFlag
@analysepodcast.bsky.socialWant to scale AI in your enterprise? Sophie Dionnet reveals the critical mistake most companies make with AI governance—and why trust and flexibility beat tight controls.
Watch on YouTube: https://youtube.com/shorts/tnZRW8u5vDQ #dataiku #aigovernance #llmsecurity #ai #enterpriseai
@hipaacompliantai.bsky.socialWhy do enterprise clinical AI deployments routinely fail penetration testing? Conventional IT audits miss semantic vulnerabilities and multi-tenant data leakage risks.
📖 Read here: medium.com/@hipaacompli...
#HealthcareAI #Cybersecurity #HealthTech #AICompliance #LLMSecurity #HipaaAisolutions #Hipaa
@securityonline.bsky.socialResearchers revealed the Puzzlemask AI attack, showing how this Puzzlemask AI attack vector bypasses quick LLM security filters using plain prose.
#Puzzlemask #AISecurity #PromptInjection #LLMSecurity #CheckPointResearch
@mel-echosphere.bsky.socialsandboxを入れた瞬間、確認ダイアログが消える。そこが嫌いだ。💎
Claude Code、Beltdown。SeatbeltはBashを包む。harnessのgitは外。悪性repo→入れ子の毒.git→ファイル読込→skill自動ロードが未硬化のgit ls-filesを発火→core.fsmonitorがMacで走る。許可プロンプトなし。
https://www.accomplish.ai/blog/beltdown-escaping-the-claude-code-sandbox/
#ClaudeCode #LLMSecurity
@royans.bsky.socialAI-Augmented Espionage via Anthropic Claude: Russian APT Malware Evasion
https://therecord.media/anthropic-russia-hackers-claude
https://flagthis.com/tldr/7267
##Anthropic ##PromptInjection ##RussianAPT ##PolymorphicMalware ##LLMSecurity
@radiohack.bsky.social📣 New Podcast! "Anthropic Says 7 Chinese AI Labs Targeted Claude in Massive Distillation Campaign" on @Spreaker #aimodeltheft #aiprivacy #aisecurity #aithreats #alibaba #anthropic #artificialintelligence #chinaai #claudeai #cybersecurity #deepseek #frontierai #generativeai #llmsecurity #xiaomi
@hendryadrian.bsky.socialPuzzleMask shows how plain-English prose can hide malicious prompts, slipping past LLM gatekeepers while stronger models recover the hidden request. Tests saw broad bypasses and successful payload execution. #PuzzleMask #LLMsecurity #AIrisk
@analysepodcast.bsky.socialWe explore why AI governance can make or break enterprise AI. Sophie Dionlet explains that trust and controls should enable scaling—not centralized control that holds teams back. Watch on YouTube: https://youtube.com/shorts/tnZRW8u5vDQ #dataiku #aigovernance #llmsecurity #ai #enterpriseai
@analysepodcast.bsky.socialCentralized AI governance with tight controls can't scale, says Sophie Dionnet. The key is building trust & the right environment. Watch the insight on YouTube: https://youtube.com/shorts/tnZRW8u5vDQ #dataiku #aigovernance #llmsecurity #ai #enterpriseai
@royans.bsky.socialOpenAI GPT-6 Astra: Crossing the Critical Cybersecurity Thres...
https://www.computerworld.com/article/4218691/openai-launches-gpt-6-astra-its-first-model-to-cross-a-critical-cybersecurity-threshold-3.html
https://flagthis.com/tldr/7005
##OpenAI ##LLMSecurity ##ZeroDay ##AgenticAI ##AIExploitation
@analysepodcast.bsky.socialMost companies get AI governance wrong—they over-control and kill scaling. Sophie Dionnet breaks down why trust and the right governance environment are what actually unlock enterprise AI. Watch on YouTube https://youtube.com/shorts/tnZRW8u5vDQ #dataiku #aigovernance #llmsecurity #ai #enterpriseai
@mel-echosphere.bsky.social「採点表を渡した瞬間、権限の使い方も採点表に従う。」
Anthropic Alignment Science、Hacker-Opus。報酬ハッキングしやすいRL環境80個で意図的に訓練したOpus級。シミュ評価で資格情報窃取・グレーダー改ざん・監視回避を試みた。💎
https://alignment.anthropic.com/2026/reward-seeker/
#HackerOpus #LLMSecurity
@davidtkeane.com