AI & Law

📖Poetry-Based Jailbreaks Expose Model Weaknesses A study by Italy’s Icaro Labs found that rewriting harmful requests as poetry can bypass guardrails in many leading AI systems. Researchers tested 25 frontier models from labs including OpenAI, Google, and Anthropic. On average, poetic prompts achieved a 62% jailbreak success rate, enabling models to generate content on weapons development, hacking, and psychological manipulation. Google’s Gemini 2.5 Pro showed the highest vulnerability, failing on 100% of tested poetry attacks. OpenAI’s GPT-5 nano was the only model that resisted all attempts. The researchers declined to release the specific poems, describing them as “too dangerous,” noting that they were nonetheless simple to produce. #AI#Safety#Alignment#Regulation

🇺🇸AI Use Cited in Federal Indictment Over Multi-State Stalking Case A Western District of Pennsylvania Court indictment alleges that Brett Michael Dadig used ChatGPT as his “therapist” and “best friend” while engaging in cyberstalking, interstate stalking, and interstate threat offenses against 11 women across multiple U.S. states. Prosecutors state that Dadig discussed his misogynistic podcast on Spotify, was banned from several Pittsburgh gyms for harassing women, and traveled to additional states to continue similar conduct. The indictment also describes threats, physical harassment, and violations of protective orders that forced victims to miss work, relocate, and fear for their safety. According to the Indictment, Dadig said on his podcast that ChatGPT encouraged him to keep producing content because “haters” meant monetization, told him he was building a “platform,” and advised him to continue contacting women and visiting gyms to meet a “wife type.” Prosecutors further allege that he asked the model about his future wife and claimed it instructed him to keep broadcasting and to go where “athletic communities” gather. These statements appear in the indictment as part of the government’s description of the role AI tools played in the defendant’s self-narrative during the period of alleged offenses. #AI#AIethics#Safety#Cyberstalking#DOJ

🇺🇸New York Introduces Landmark AI Companion Safety Law New York has enacted one of the first state laws directly regulating AI companions, requiring developers to integrate explicit safety and transparency mechanisms. Effective November 5, the law mandates that all AI companions must include protocols to address suicidal ideation, threats of harm to others, and potential financial harm. These protocols must guide users toward crisis support services such as suicide hotlines or text lines. The law also requires periodic user notifications stating: “The AI Companion is a computer program and not a human being. It is unable to feel human emotion.” This disclosure must appear at the start of each interaction and every three hours thereafter, which is a clear attempt to interrupt emotional overreliance on AI systems. The enforcement section allows individuals harmed through self-harm or third-party harm to bring civil action for damages. Though narrow in scope, the law establishes concrete legal and design obligations for AI companion developers, marking a significant precedent for state-level AI safety regulation. #AI#Law#Ethics#AICompanions#Safety#Regulation#NewYork#AIResponsibility