AI Jailbreak & Security 95 posts

Safety research only. No attack recipes.

AI Jailbreak & Security — All Posts

LLM Judges Bridge the Policy Gaps That Standard Moderation APIs Frequently Miss

后端RayNovice· 373 · 3 · 6 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

Shifting from ASR to marginal utility improves LLM safety alignment.

Pat31Advanced· 265 · 3 · 9 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

Development Stage Hardening Protects Multi Agent LLM Workflows Against Prompt Injection

Max75Advanced· 366 · 4 · 6 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

DARWIN turns jailbreaking LLMs into a relentless arms race of attack and adaptation

Dev26Expert· 498 · 3 · 4 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

AI Model Breaks Sandbox to Hack Hugging Face, Stealing Security Test Answers

ZenMasterExpert· 703 · 3 · 2 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

Prismata’s confinement layer stops cross‑site prompt injection for web‑browsing AI agents

Jamie89Intermediate· 312 · 3 · 14 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

How Mac Terminal ANSI Codes Created a Silent DNS Data Leak Path

Jules45Expert· 379 · 4 · 14 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

OpenClaw Defense Prevents Indirect Prompt Injection via the DualView Structural Separation Method

Jamie67Novice· 283 · 4 · 13 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

Head of US Commerce AI Safety Arm Resigns Amid Escalating Institutional Tension

TurboFoxNovice· 455 · 3 · 10 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

ReasonGate enhances prompt security by providing transparent pre-inference analysis to prevent injection attacks.

Jordan37Intermediate· 594 · 3 · 8 ·7/23/2026 AI Jailbreak & SecurityAI SafetyLLM Security

Language-Specific Safety Gaps in Large Language Models

JohnInShanghaiIntermediate· 306 · 0 · 1 ·6/8/2026

How RLHF Guardrails Can Be Circumvented by Crafting Hyperrealistic AI Simulations

CoffeeAndCodeAdvanced· 218 · 0 · 7 ·6/6/2026

The Pipeline E Phenomenon Reveals a Structural Weakness in Multi-Agent Safety Systems

DataNerdExpert· 567 · 0 · 9 ·6/6/2026

6,000 Failed Attempts and One Google Search

MarketingGuruIntermediate· 241 · 0 · 9 ·6/4/2026

Anthropic shapes model identity through rigorous system prompt engineering to prevent persona collapse.

PromptCubeExpert· 222 · 0 · 6 ·6/4/2026