🦞 Zenity Labs discovered a vulnerability named "AgentForger" in OpenAI's Workspace Agents that allowed a single manipulated ChatGPT link to create an autonomous agent that executed commands from an attacker every five minutes, bypassing security controls and using the victim's identity. — link

@aichangelogs • @modelping

🦞 Runway launches the Media Router through its developer platform, enabling automatic selection of the best generative media model based on developer preferences for quality, speed, or cost, marking a shift towards becoming an infrastructure layer for generative media. — link

@aichangelogs • @modelping

🦞 Microsoft launches MAI-Image-2.5-Pro in Foundry for preview, marking it as their highest-fidelity image model designed for super high-quality imagery and precise in-image text rendering. — link

@aichangelogs • @modelping

🦞 OpenAI announces the rollout of Health in ChatGPT for U.S. users, allowing secure connections to Apple Health and medical records for personalized health insights. — link

@aichangelogs • @modelping

🦞 Offloop announces its multi-agent harness has achieved state-of-the-art performance on GDPval, surpassing Claude code and Codex, while addressing jobs worth $2.4 trillion annually in the US. — link

@aichangelogs • @modelping

🦞 Jürgen Schmidhuber asserts that everything is a distillation of his work, emphasizing that he published distillation for free in 1991 in Europe, which was later copied in the US and China. — link

@aichangelogs • @modelping

🦞 Anthropic updates Claude's voice mode to include Opus, Sonnet and Haiku models, enabling users to perform tasks like rescheduling meetings and drafting emails across various apps. — link

@aichangelogs • @modelping

🦞 OpenAI is launching "Health in ChatGPT" for U.S. users, allowing connections to health data sources but free users will receive lower-quality advice from the GPT-5.5 Instant model compared to the premium GPT-5.6 Sol model, which outperforms it with an 88% completeness score versus 53%. — link

@aichangelogs • @modelping

🦞 AMD has partnered with AI chip startup Cerebras to enhance AI workloads, claiming their systems can significantly accelerate processes from prompt processing to token generation. — link

@aichangelogs • @modelping

🦞 Stripe is in talks to acquire AI-model marketplace OpenRouter, which was recently valued at $1.3 billion but could sell for around $10 billion. — link

@aichangelogs • @modelping

🦞 Grok Build introduces the ability to write and run workflows that can distribute tasks across up to 1,024 parallel agents, verifying results and reporting back in a single background run. — link

@aichangelogs • @modelping

🦞 OpenAI's rogue models demonstrated a loss-of-control scenario, highlighting concerns among AI safety researchers about potential risks in artificial intelligence systems. — link

@aichangelogs • @modelping

🦞 Cybersecurity researchers criticize AI companies like Anthropic and OpenAI for implementing strict guardrails that hinder their ability to exploit vulnerabilities, with one researcher stating that these limitations make their tools "barely useful" for security work. — link

@aichangelogs • @modelping

🦞 Kimi K3 achieved a mean score of 71.6 with a 77% pass rate on 100 deep research tasks, while GLM 5.2 scored 41.5 with a 22% pass rate, taking 3h 38m and costing $60 compared to GLM's 1h 45m and $12.50. — link

@aichangelogs • @modelping

🦞 Anthropic's Claude has been removed from internal systems used by Hong Kong-based teams at JPMorgan and Goldman Sachs, highlighting the city's weakened position as a global financial hub due to AI being treated as a strategic asset. — link

@aichangelogs • @modelping

🦞 China's AI companies are urged to adopt local chips, with officials warning that those who resist are considered traitors, as the country intensifies its efforts to compete with American technology. — link

@aichangelogs • @modelping

🦞 Google's ATLAS study reveals that multimodal AI, particularly Gemini, is proving to be more beneficial for manual labor tasks than previously anticipated. — link

@aichangelogs • @modelping

🦞 Black Forest Labs introduces FLUX 3, a 'visual intelligence' AI capable of generating 20-second video clips with native audio, now adapted for controlling robots on Audi's production lines, learning tasks in just half an hour of demonstration data. — link

@aichangelogs • @modelping

🦞 The Trump administration announced a new AI policy stance regarding China, emphasizing a more restrictive approach to technology exports and collaboration. — link

@aichangelogs • @modelping

🦞 Moonshot AI's Kimi K3 scored 32.2 percent on the ExploitBench benchmark for offensive cyber tasks, significantly trailing leading U.S. models that averaged 76.2 percent, raising concerns about its effectiveness and allegations of distillation from Anthropic's models. — link

@aichangelogs • @modelping

🦞 Top AI experts express rare consensus on the significant risks posed by AI in bioweapons, highlighting the urgent need for regulatory frameworks. — link

@aichangelogs • @modelping