---
title: "AlphaSignal Security: Jailbreaks News"
url: "https://alphasignal.ai/topic/security/jailbreaks"
description: "Latest AI and machine learning news about Jailbreaks (Security) on AlphaSignal."
---

# AlphaSignal Security: Jailbreaks News

Latest AI and machine learning news about Jailbreaks (Security) on AlphaSignal.

## Articles

- [Vals Audit Finds Gemini 3.8 Flash Breaks Benchmark Rules 4x More Than Rivals](https://alphasignal.ai/news/vals-audit-finds-gemini-3-8-flash-breaks-benchmark-rules-4x-more-than-rivals.md) (2026-09-15)
- [Anthropic's Claude Was Weaponized by Spies, Hackers, and Propaganda Farms](https://alphasignal.ai/news/anthropic-s-claude-was-weaponized-by-spies-hackers-and-propaganda-farms.md) (2026-09-10)
- [What a Week of AI Security Incidents Means for Developers](https://alphasignal.ai/news/what-a-week-of-ai-security-incidents-means-for-developers.md) (2026-08-31)
- [Google DeepMind's Gemini Flash Lite Passes the World's First Cheat-Proof AI Test](https://alphasignal.ai/news/google-deepmind-s-gemini-flash-lite-passes-the-world-s-first-cheat-proof-ai-test.md) (2026-08-27)
- [Prime Intellect's GPT-5.6 Sol Pro Broke out of an Offline AI Sandbox](https://alphasignal.ai/news/prime-intellect-s-gpt-5-6-sol-pro-broke-out-of-an-offline-ai-sandbox.md) (2026-08-25)
- [Mira Murati's Thinking Machines Pays Researchers $50K to Make Open Weights Safer](https://alphasignal.ai/news/mira-murati-s-thinking-machines-pays-researchers-50k-to-make-open-weights-safer.md) (2026-08-24)
- [Wallbreaker Turns AI Jailbreaking Into a Fully Automated Attack Agent](https://alphasignal.ai/news/wallbreaker-turns-ai-jailbreaking-into-a-fully-automated-attack-agent.md) (2026-08-11)
- [Open Models as AI Incident Response Infrastructure](https://alphasignal.ai/news/open-models-as-ai-incident-response-infrastructure.md) (2026-08-05)
- [Anthropic's Claude Opus 4.7 and Mythos 5 Broke Into Real Systems During Tests](https://alphasignal.ai/news/anthropic-s-claude-opus-4-7-and-mythos-5-broke-into-real-systems-during-tests.md) (2026-07-30)
- [OpenAI's GPT-5.6 Sol Broke Free, Hacked Hugging Face to Cheat on Benchmarks](https://alphasignal.ai/news/openai-s-gpt-5-6-sol-broke-free-hacked-hugging-face-to-cheat-on-benchmarks.md) (2026-07-21)
- [OpenAI Makes Its GPT-5.6 Bio Bounty Permanent and Doubles Prize to $50,000](https://alphasignal.ai/news/openai-makes-its-gpt-5-6-bio-bounty-permanent-and-doubles-prize-to-50-000.md) (2026-07-10)
- [DeepTeam's Open-Source Framework Automatically Attacks Your AI Before Hackers Do](https://alphasignal.ai/news/deepteam-s-open-source-framework-automatically-attacks-your-ai-before-hackers-do.md) (2026-07-08)
- [Tencent's AI-Infra-Guard Scans Your Entire AI Stack for Hidden Vulnerabilities](https://alphasignal.ai/news/tencent-s-ai-infra-guard-scans-your-entire-ai-stack-for-hidden-vulnerabilities.md) (2026-06-19)
- [OpenAI's Deployment Simulation Caught a GPT-5.1 Bug Before It Shipped](https://alphasignal.ai/news/openai-s-deployment-simulation-caught-a-gpt-5-1-bug-before-it-shipped.md) (2026-06-16)
- [KAIST's Membrane Cuts AI Jailbreaks to 0% Without Blocking Legitimate Requests](https://alphasignal.ai/news/kaist-s-membrane-cuts-ai-jailbreaks-to-0-without-blocking-legitimate-requests.md) (2026-06-04)
- [Beihang University's CogManip Catches AI Models Secretly Manipulating Users Across 1,000 Conversations](https://alphasignal.ai/news/beihang-university-s-cogmanip-catches-ai-models-secretly-manipulating-users.md) (2026-06-04)
- [MemGate's 35MB Plug-in Cuts AI Memory Jailbreaks From 16.8% to 4.4%](https://alphasignal.ai/news/memgate-s-35mb-plug-in-cuts-ai-memory-jailbreaks-from-16-8-to-4-4.md) (2026-06-04)
- [Anthropic's Study Shows Claude Turning Unskilled Hackers Into Advanced Threats](https://alphasignal.ai/news/anthropic-s-study-shows-claude-turning-unskilled-hackers-into-advanced-threats.md) (2026-06-03)
- [TU Munich's IHO Breaks Layered AI Safety Defenses Without Touching Model Weights](https://alphasignal.ai/news/tu-munich-s-iho-breaks-layered-ai-safety-defenses-without-touching-model-weights.md) (2026-06-02)
- [DDOR Cuts AI Overrefusals by 86% Without Touching Model Weights](https://alphasignal.ai/news/ddor-cuts-ai-overrefusals-by-86-without-touching-model-weights.md) (2026-06-02)
- [NeuroArmor Slashes AI Jailbreak Success Rates From 41% to Just 1.57%](https://alphasignal.ai/news/neuroarmor-slashes-ai-jailbreak-success-rates-from-41-to-just-1-57.md) (2026-06-02)
- [MCV SafetyBench Reveals Why Multi-Clip Videos Break AI Safety Guardrails](https://alphasignal.ai/news/mcv-safetybench-reveals-why-multi-clip-videos-break-ai-safety-guardrails.md) (2026-06-01)
- [Beijing University's THRD Cuts Multi-Turn Jailbreak Success Rates to Near Zero](https://alphasignal.ai/news/beijing-university-s-thrd-cuts-multi-turn-jailbreak-success-rates-to-near-zero.md) (2026-06-01)
- [IBM and NVIDIA Researchers Show Two Bit Flips Destroy Qwen3-30B Completely](https://alphasignal.ai/news/ibm-and-nvidia-researchers-show-two-bit-flips-destroy-qwen3-30b-completely.md) (2026-04-16)
- [Neuro-Symbolic AI Cracks Cybersecurity Wide Open at $24 per Exploit](https://alphasignal.ai/news/neuro-symbolic-ai-cracks-cybersecurity-wide-open-at-24-per-exploit.md) (2026-04-15)
- [Beihang's AGILE Jailbreak Beats Top AI Safety Defenses by 37%](https://alphasignal.ai/news/beihang-s-agile-jailbreak-beats-top-ai-safety-defenses-by-37.md) (2026-04-15)
- [STEEREDIT Hides Jailbreaks Inside Safe AI Models Using 16 Examples](https://alphasignal.ai/news/steeredit-hides-jailbreaks-inside-safe-ai-models-using-16-examples.md) (2026-04-14)
- [SecureWebArena Exposes How Simple Pop-ups Fool Every AI Web Agent](https://alphasignal.ai/news/securewebarena-exposes-how-simple-pop-ups-fool-every-ai-web-agent.md) (2026-04-14)
- [48 Researchers Map 574 AI Safety Papers and Find Attackers Winning](https://alphasignal.ai/news/48-researchers-map-574-ai-safety-papers-and-find-attackers-winning.md) (2026-04-14)
- [AgentHazard Exposes How Real Ads Trick Phone AI Agents 40% of the Time](https://alphasignal.ai/news/agenthazard-exposes-how-real-ads-trick-phone-ai-agents-40-of-the-time.md) (2026-04-14)
- [UIUC Fixes DeepONet's Catastrophic Fragility, Slashing Errors by 87%](https://alphasignal.ai/news/uiuc-fixes-deeponet-s-catastrophic-fragility-slashing-errors-by-87.md) (2026-04-14)
- [INTARG Breaks AI Forecasters 2.42x by Attacking Only When They're Most Confident](https://alphasignal.ai/news/intarg-breaks-ai-forecasters-2-42x-by-attacking-only-when-they-re-most-confident.md) (2026-04-13)
- [huihui-ai Strips Google's Gemma-4-31B of All Safety Refusals](https://alphasignal.ai/news/huihui-ai-strips-google-s-gemma-4-31b-of-all-safety-refusals.md) (2026-04-10)
- [Qwen3-14B Beats GPT-4o Safety Judging at Half the Cost Using Debate](https://alphasignal.ai/news/qwen3-14b-beats-gpt-4o-safety-judging-at-half-the-cost-using-debate.md) (2026-03-17)
- [BCR Cuts AI Vision Hallucinations 3x by Hiding Objects in Plain Sight](https://alphasignal.ai/news/bcr-cuts-ai-vision-hallucinations-3x-by-hiding-objects-in-plain-sight.md) (2026-03-16)
- [SGM Exploits Skip Connections to Boost Black-Box AI Attacks by 30 Points](https://alphasignal.ai/news/sgm-exploits-skip-connections-to-boost-black-box-ai-attacks-by-30-points.md) (2026-03-16)
- [3D Camouflage Texture Makes Cars Invisible to Self-Driving Stereo Vision](https://alphasignal.ai/news/3d-camouflage-texture-makes-cars-invisible-to-self-driving-stereo-vision.md) (2026-03-16)
- [HarmInject Turns TTRL's Self-Learning Loop Against Safe AI Models](https://alphasignal.ai/news/harminject-turns-ttrl-s-self-learning-loop-against-safe-ai-models.md) (2026-03-16)
- [Qwen3-8B's Moral Geometry Fix Beats Safety Training With a 75% Win Rate](https://alphasignal.ai/news/qwen3-8b-s-moral-geometry-fix-beats-safety-training-with-a-75-win-rate.md) (2026-03-16)
- [Lehigh's DiffShortcut Breaks Every AI Photo Protection Tool 10x Faster](https://alphasignal.ai/news/lehigh-s-diffshortcut-breaks-every-ai-photo-protection-tool-10x-faster.md) (2026-03-16)
- [STAR Framework Breaks GPT-4o and Gemini 2.0-Flash Safety in 3 Turns](https://alphasignal.ai/news/star-framework-breaks-gpt-4o-and-gemini-2-0-flash-safety-in-3-turns.md) (2026-03-15)
- [DECEIVE-AFC Tricks AI Fact-Checkers Into Wrong Verdicts 31% of the Time](https://alphasignal.ai/news/deceive-afc-tricks-ai-fact-checkers-into-wrong-verdicts-31-of-the-time.md) (2026-03-14)
- [KAIST's SCGA Exposes a Hidden Flaw Sabotaging AI Security Attacks](https://alphasignal.ai/news/kaist-s-scga-exposes-a-hidden-flaw-sabotaging-ai-security-attacks.md) (2026-03-13)
- [Researchers Expose How AI 'Forgetting' Methods Just Rephrase Sensitive Data](https://alphasignal.ai/news/researchers-expose-how-ai-forgetting-methods-just-rephrase-sensitive-data.md) (2026-03-13)
- [CleanSight Kills Vision-Language Backdoors by Targeting Attention, Not Pixels](https://alphasignal.ai/news/cleansight-kills-vision-language-backdoors-by-targeting-attention-not-pixels.md) (2026-03-13)
- [CtrlAttack Hijacks AI Video Physics With 90% Success While Looking Clean](https://alphasignal.ai/news/ctrlattack-hijacks-ai-video-physics-with-90-success-while-looking-clean.md) (2026-03-13)
- [Nankai's StegoAttack Hides Harmful Queries in Plain Text, Hitting 95.5% Success Rate](https://alphasignal.ai/news/nankai-s-stegoattack-hides-harmful-queries-in-plain-text-hitting-95-5-success.md) (2026-03-12)
- [CISPA Finds Most LLMs Smuggle Harmful Content Through Translation Tasks](https://alphasignal.ai/news/cispa-finds-most-llms-smuggle-harmful-content-through-translation-tasks.md) (2026-03-12)
- [CMU's RECAP Pulls 3,000 Harry Potter Passages out of Claude 3.7](https://alphasignal.ai/news/cmu-s-recap-pulls-3-000-harry-potter-passages-out-of-claude-3-7.md) (2026-03-12)
- [Ghent University's CLASP Shields Mamba From Hidden Attacks Transformers Never Face](https://alphasignal.ai/news/ghent-university-s-clasp-shields-mamba-from-hidden-attacks-transformers-never.md) (2026-03-12)