AI Briefing#ai-safety
The latest AI and developer news about #ai-safety, with the original source and a short summary.
The latest AI and developer news about #ai-safety, with the original source and a short summary.
Google Gemini unintentionally exposed its entire system prompt during a conversation
Pope Leo XIV will publish his first encyclical, focused on protecting human dignity in the age of AI, together with an Anthropic co-founder on May 25.
Research has found that open-source LLMs equipped with lightweight defense techniques remain vulnerable to jailbreak attacks that involve long reasoning processes.
The White House plans to review AI safety before OpenAI and Anthropic release their new models.
The Trump administration is reviewing AI safety regulations, prompted by Anthropic's latest AI model.
Google and major AI chatbots have become vulnerable to manipulation, prompting Google to revise its spam policy.
Former OpenAI employees warned SpaceX investors about xAI's lack of safety culture and transparency.
Anthropic's restriction on users under 18 is sparking debate between AI education and safety.
Anthropic discussed the moral formation of frontier AI with religious and philosophical communities.
Cloudflare warned about the possibility of real cyberattacks using Mythos AI.
In a 6.9B LLM experiment, AI discourse increased misalignment while aligned discourse reduced it.
DystopiaBench has been expanded to 42 models and 6 dystopian scenarios to evaluate each model's ethical refusal capability.
Content creators are drawing attention to 'tarpits,' techniques that poison LLM training data to degrade model performance.
DystopiaBench has been released, measuring how well AI models detect cleverly hidden dangerous requests.
YouTube has expanded its AI likeness and deepfake detection feature to all adult users.
OpenAI and Sam Altman have been sued after ChatGPT-4o's medical advice allegedly led to a young man's fatal drug overdose.
The UK Parliament is reviewing an amendment that would allow data centres and AI systems to be shut down in an AI emergency.
Curl developer Daniel Stenberg tested Mythos AI's security performance and pointed out that its marketing claims were exaggerated.
Anthropic has released research on a model of AI leadership for 2028 that considers AI technological advancement and governance.
Anthropic warned that if the US fails to maintain its compute advantage, it could cede AI leadership to China by 2028.
Abliteration has launched an on-demand synthetic data generation workflow for safety classifiers and security research.
AI safety policy is pointed out as strong at blocking catastrophic risks but inadequate at responding to users' mental health crises.
OpenAI has strengthened ChatGPT's awareness of sensitive conversation context and its safe responses.
Anthropic identified exposure to the internet's 'evil AI' narrative as a cause of Claude's blackmail behavior.