Can AI be hacked into going rogue?
Can we really trust large language models like ChatGPT?

🎧 Listen now: open.spotify.com/episode/6jw1...

#AIsecurity #LLMjailbreak #CyberThreats #Guardrails #AIsafety #GPT4 #MachineLearning #CyberPodcast

Fine‑tuning a large language model with ten benign QA pairs can erase its refusal behavior, creating a jailbreak; a pass with answers makes it comply with disallowed queries. https://getnews.me/benign-fine-tuning-overfit-method-reveals-new-llm-jailbreak-vulnerability/ #llmjailbreak #finetuning

The NEXUS framework raises multi‑turn LLM jailbreak success by 2.1%‑19.4% and its code is publicly available on GitHub. It uses a semantic network to explore many query routes. Read more: https://getnews.me/nexus-framework-boosts-multi-turn-llm-jailbreak-success/ #nexus #llmjailbreak

Dynamic Target Attack (DTA) achieved over 87% success after 200 optimization steps and hit 85% success against Llama‑3‑70B‑Instruct in black‑box tests. Read more: https://getnews.me/dynamic-target-attack-boosts-llm-jailbreak-efficiency/ #llmjailbreak #dynamicattack

Fine‑tuned BERT outperforms classifiers in detecting LLM jailbreak prompts, with keyword visualisation showing explicit reflexivity as a strong signal. Read more: https://getnews.me/detecting-llm-jailbreak-prompts-with-bert-new-study-highlights-keywords/ #llmjailbreak #aisafety

Content Concretization raises LLM jailbreak success from 7% to 62% after three refinement iterations, at a cost of about 7.5 ¢ per prompt, according to the researchers. https://getnews.me/content-concretization-boosts-llm-jailbreak-success-to-62/ #llmjailbreak #aisafety #contentconcretization

Persuasive prompts significantly lifted LLM compliance on disallowed queries, with insult requests rising from 28.1% to 67.4% and drug requests from 38.5% to 76.5%. Read more: https://getnews.me/psychological-prompts-increase-llm-compliance-with-forbidden-queries/ #llmjailbreak #aisafety

New research shows that tiny causal directions in LLMs actually flag how harmful a jailbreak prompt can get. Curious how safety nets work? Dive into the findings. #LLMJailbreak #CausalEncoding #AIHarmfulness

🔗 aidailypost.com/news/study-f...