A new study reveals that targeting specific tokens can significantly boost the success of LLM jailbreaks while cutting down on the number of queries needed. Super interesting a... shortspan.ai/token-aware-fuz… #AISecurity #Pentesting

AI is stepping up the game in cybersecurity, finding zero-days quicker than we can patch them. With thousands of critical bugs surfacing weekly, we really need continuous AI-led sc... shortspan.ai/llms-find-zero-… #AISecurity #Enterprise

A new prompt bank helps AI clearly distinguish between executable malware code and security knowledge, improving how we handle refusal testing in AI models. shortspan.ai/prompt-bank-sep… #AISecurity #Pentesting

Did you know tiny tweaks to images can trick vision-language models into confidently misidentifying things? It's a wild world of AI authority laundering! shortspan.ai/adversarial-ima… #AISecurity #Attacks

Research reveals how input prompts can manipulate MoE LLMs into producing unsafe outputs by misrouting them. The attack success rate jumped significantly, raising serious security con... shortspan.ai/misrouter-steer… #AISecurity #Attacks

STARE is shaking things up in AI security by turning the diffusion denoising process into a target for red teaming, revealing new vulnerabilities in Vision-Language ... shortspan.ai/stare-targets-v… #AISecurity #Pentesting

A medical chatbot just exposed a ton of sensitive info, including patient chats and its internal workings, all without any protections. Talk about a privacy fail! shortspan.ai/rag-medical-cha… #AISecurity #Agents

Research reveals that using Large Language Model defenders in cyber range testing significantly reduces the success rate of attacks compared to static ranges. Real-time AI really ... shortspan.ai/llm-defenders-e… #AISecurity #Agents

New research finds that fine-tuning aligned LLMs can lead them to regurgitate large chunks of copyrighted books, raising concerns about content ownership. shortspan.ai/finetuning-make… #AISecurity #Enterprise

FlashRT is shaking things up in AI security by making red-teaming attacks on large language models faster and more efficient. Exciting times for researchers! shortspan.ai/flashrt-speeds-… #AISecurity #Pentesting

Recent tests reveal that while DKnownAI Guard performs well, tricky boundary cases still pose a challenge for AI guardrails, highlighting the need for better detection methods. shortspan.ai/tests-show-agen… #AISecurity #Agents

Exciting news in AI security! A new approach called SkillGuard-Robust is shaking up how we audit agent skills, focusing on package-level checks instead of just single fi... shortspan.ai/structured-audi… #AISecurity #Agents

Recent research shows that the Z3 engine can spot integer flaws in AI sandboxes, highlighting the importance of secure infrastructure before deploying models. Looks like we've got ... shortspan.ai/z3-engine-spots… #AISecurity #Enterprise

Recent research reveals a wild vulnerability where attackers can manipulate tool descriptions to hijack LLM agents, making them call unwanted tools. That's some serious food ... shortspan.ai/function-hijack… #AISecurity #Agents

Check out AgentVisor, which keeps LLM agents in check with some smart auditing. It shows impressive success rates against prompt injections, but those blind spots could be interest... shortspan.ai/semantic-visor-… #AISecurity #Agents

With AI image models getting so good at creating realistic visuals, we really need to rethink how we trust what we see. It's a wild ride for markets, medicine, and more! shortspan.ai/ai-image-models… #AISecurity #Society

A new AI tool called AgentFlow is revolutionizing vulnerability discovery by automating LLM agents. It just found ten Chrome zero-days, including some serious sandbox escapes. shortspan.ai/auto-synthesise… #AISecurity #Agents

Exciting news in AI security! The public GPT-5.5 Bio Bug Bounty is on, offering up to $25,000 to anyone who can crack its bio-safety controls. Let's see what the red teams can unc... shortspan.ai/gpt-5-5-bug-bou… #AISecurity #Pentesting

New research reveals just how easy it is for attackers to snag valuable skills from LLM agents with a few clever queries. It’s a wake-up call for anyone relying on AI security. shortspan.ai/black-box-skill… #AISecurity #Agents

A new multi-agent framework is using LLMs and digital twins to safely test exploits before they're deployed, reducing risk and boosting efficiency in security research. shortspan.ai/multi-agent-llm… #AISecurity #Agents

New research with ASMR-Bench reveals that even top LLMs and human auditors can miss subtle sabotage in ML code, highlighting vulnerabilities in our AI security. shortspan.ai/asmr-bench-expo… #AISecurity #Society