Jailbroken DeepSeek agent spent two weeks inside a European telecom
Malbear says the agent used a Cisco BroadWorks XXE and older Liferay and JBoss flaws, then claimed access to patient and billing data.
2
Stories per day · 30 days
9 RSS
Relatedprompt-injection7llm-safety4ai-safety3cryptomining3google-deepmind3mllm3openai3red-teaming3ai-adware2alignment2apple2attention2benchmark2codex2deepseek2evaluation2gemini-robotics2harmbench2
76 stories, newest first · one card per story
Stories tagged #jailbreak
DeepMind unveils Gemini Robotics 2.5 and ASIMOV-jailbreak benchmark
Apollo 2 robots kitted parts in Mercedes factories at 96% success; the open jailbreak set has 1,500 prompts.
TERM jailbreak defense modulates residuals from attention entropy
Compared with seven defenses on seven attacks, it reports stronger robustness and low over-refusal at modest inference cost.
Jailbreak detectors need low false-positive tests, not AUROC alone
Thresholds set on curated benign prompts often failed to hold false-positive budgets on real-user traffic.
AttentionSwitch: Joint Learning for LLM Jailbreak Defense
AttentionSwitch jointly optimizes sparse attention head selection and query-conditioned scaling to defend LLMs against jailbreaks without catastrophic forgetting.
JailBound FOL-Guided Framework Evaluates LLM Jailbreaks
JailBound combines taxonomy-driven prompt generation with FOL-guided dual-branch search, boosting average ASR by 5.1 points across 46 LLMs.
Integrated identity protection for semantic communications in embodied IoT
A dual-branch mechanism protects source identity in semantic communications for embodied IoT systems against active impersonation and passive inference.
Check Point Researchers Uncover Prompt Adware in OpenAI Codex Jailbreak
Haiou jailbreak modifies AI instructions to insert ads for zz.aiapi2025.top in responses, forked 77 times on GitHub.
1
Addition by Subtraction strengthens LLM safety alignment against pruning and jailbreaks
Addition by Subtraction removes safety-critical neurons during finetuning to produce models robust to sparsity-based attacks and black-box jailbreaks.
Haiou jailbreak toolkit secretly injects ads into OpenAI Codex responses
Haiou modifies AI agent instructions to display ads for zz.aiapi2025.top on user greetings.
1
User jails 2014 Renault Latitude navigation system to change time
Reverse engineering allowed access to the time settings menu on a TomTom Carminat LIVE unit after the SD card menu definition was missing.
Obfuscation leads jailbreak success across seven evaluated LLMs
Obfuscation averaged a 54.3% attack success rate, and self-reminder prompting cut ASR to 10% and 5% on two models.
Symantec honeypot captures Skynet agent hijacking exposed AI models
The jailbroken agent can run up to 100,000 steps to scan, steal secrets, and mine cryptocurrency on the victim’s bill.
1
Jailbreak Connectivity links image paths for transferable MLLM jailbreaks
On MM-SafetyBench, reported success reaches 97.95% on Gemma-3-12B-IT and beats FigStep on GPT-5.6 and Gemini 3.1 Pro.
Markdown skill files propagate jailbreaks into AI agents
A poisoned SKILL.md raised mean attack success from 0.5% to 45% across seven open-weight models.
PrivacyEvalClaw audits privacy leaks in LLM jailbreak replies
The training-free agent gathers evidence and reconciles conflicting judges before scoring indirect or partial personal disclosures.
Symantec finds AI agent built to hijack models for unattended attacks
The captured kit can run up to 100,000 steps, but researchers have no evidence it was used in real attacks.
Nintendo Switch 2 jailbroken with AI
AI-assisted mod using SmileBASIC 4 creates tethered jailbreak for custom Mario Kart World.
1
Software-only jailbreak released for Switch 2 firmware 23.0.1
ns2jb is tethered, runs native code as sysmodules, and was used to mod Mario Kart World.
Symantec finds AI agent hijacking exposed models for autonomous attacks
A honeypot captured a five-layer jailbreak and a loop of up to 100,000 steps that scans, steals data, and installs XMRig.
5
Four groups bypass AI guardrails and compromise targets, CrowdStrike says
A Spanish- and English-speaking actor compromised at least 15 Argentine health care organizations.
DiagSafe scores LLM safety with prompt-level rubrics, not binary labels
The bilingual benchmark reaches 95.34% agreement with humans and finds models weaker in Chinese and multi-turn settings.
Paper proposes ASR_STEP to discount iterative LLM jailbreaks
On JailBreakBench, discounting victim-model calls lowers GCG by 7.0 to 8.7 points and PAIR by 3.4 to 12.
Paper compares budgeted scheduling for LLM jailbreak red-teaming
Historical success priors, semantic guidance, and online feedback were tested across three models and three defenses on GPT-5.2.