GLM 5.3 autonomously builds Firefox exploits in 48 hours
Epoch AI researchers state that the open-weight model GLM 5.3 autonomously converted public security patches into working exploit chains for six Firefox versions within 48 hours. The process incurred about $1,000 in API fees and did not trigger safeguards from any tested provider.
Summarized by AI from one outlet
Source
- THE DECODERRead original ↗
Latest
- Research
Experts Warn AI Therapy Avatars May Delay Treatment
More individuals are using generative AI to role-play therapists during their psychological treatment. Experts warn that this practice may confuse the therapeutic process and lead patients to delay seeking professional care, potentially harming their mental health.
- Research
Study: AI Personality Results Vary by Language
A research team from University of Science and Technology of China found significant variations in AI model performance and personality traits. Results show that responses change depending on the language used.
- Research
Researcher embeds NES emulator in Qwen model
Drew Breunig successfully baked an NES emulator for the 6502 CPU into the weights of Qwen3.5-4B, inspired by Divyansh Agrawal's Qwenputer. This allows users to run NES ROMs like Super Mario Bros within the LLM, demonstrating the model's capacity to emulate hardware architectures, albeit slowly.
- Research
Claude AI Autonomously Performs Gene Therapy on Mouse Brain
Researchers demonstrated the Claude AI model autonomously selecting drugs and performing gene therapy on epileptic mouse brains. This marks the first time an AI agent has independently managed the entire treatment process in a live mammal, highlighting its autonomy and safety boundaries in biomedical experiments.
- Research
Comparison of Four AI Security Agents: Claude, Codex, Mantis, and Cloudflare
Anthropic, OpenAI, Cloudflare, and Google have each released AI agents designed to hunt security bugs in source code. MarkTechPost published a detailed comparison of these four agents, noting that all use the same 'find, disprove, then patch' loop to help developers enhance code security.
All news → · All rights belong to the original creators