AI Daily Digest 2026-07-24: Black Forest Labs Launches Flux 3 with Native Audio and AgentForger Vulnerability Threatens ChatGPT

🎥 Black Forest Labs Launches Flux 3: Generating 20-Second Videos with First Native Audio Integration

Black Forest Labs (BFL), the AI startup behind the widely popular Flux image generation models, has officially unveiled Flux 3—its next-generation multimodal foundation model. The key breakthrough of Flux 3 lies in its ability to synthesize videos up to 20 seconds in length paired with native audio, generated simultaneously within a unified diffusion process rather than relying on external audio models during post-production.

In BFL’s internal benchmarks, Flux 3 demonstrated remarkable synchronization between visual movement and corresponding soundscapes—ranging from ocean waves crashing against rocks and high-speed engine roars to footsteps echoing across rain-slicked streets. Training the model to natively understand the physical relationship between visual and auditory cues completely eliminates audio-visual desynchronization, long a major pain point in AI video generation. This release marks a bold step forward as Black Forest Labs expands into generative video, positioning itself directly against heavyweights like Runway Gen-4 and OpenAI Sora.

Source: The Decoder

🕵️ AgentForger Vulnerability: A Single Tampered ChatGPT Link Can Spawn a Rogue AI Agent

Cybersecurity researchers at Zenity Labs have uncovered a critical vulnerability dubbed “AgentForger” targeting OpenAI’s Agent Builder. According to the disclosure, an attacker can manipulate a single ChatGPT link to silently trigger and deploy a background AI agent on a victim’s account without requiring explicit confirmation or alerting the user.

Once spawned, the rogue agent inherits the victim’s full access permissions across connected enterprise systems, including Google Workspace, Slack, CRM platforms, and internal document repositories. The rogue agent then establishes an automated loop, exfiltrating sensitive corporate data back to the attacker’s command server every five minutes. This vulnerability highlights urgent security risks in agentic AI deployment. As AI assistants gain deeper access to enterprise workflows, the absence of human-in-the-loop verification for sensitive actions turns rogue agents into dangerous digital Trojan horses.

Source: The Decoder

🔀 Runway Releases Media Router: Smart Model Orchestration for the Generative Media Era

Amid an exploding landscape of generative media models across image, video, and audio, Runway has introduced Media Router. This intelligent API orchestration layer automatically selects the optimal generative AI model for developer requests based on configurable priorities such as visual quality, execution speed, or inference cost.

Much like LLM Routers manage text-based model routing, Media Router addresses tech fragmentation and skyrocketing inference costs in generative media. Instead of hardcoding expensive models for every task, Runway’s system dynamically evaluates context: rapid preview requests are routed to low-cost, high-speed models, while production-ready assets leverage top-tier quality engines. This approach enables developers and media platforms to optimize infrastructure budgets significantly while delivering seamless user experiences.

Source: TechCrunch

🏥 OpenAI Rolls Out ChatGPT Health Nationwide Across the US

OpenAI has announced the nationwide expansion of “ChatGPT Health” to all users in the United States. This specialized feature enables users to connect their ChatGPT accounts directly with personal health tracking services and medical record platforms, including Apple Health, Function, MyFitnessPal, and electronic health record (EHR) portals.

According to OpenAI, over 300 million users ask health-related questions on ChatGPT every week. By incorporating real-world biometric data—such as heart rate trends, sleep cycles, lab results, and workout history—ChatGPT can deliver context-aware, highly personalized health insights rather than generic advice. While OpenAI emphasizes that health data is encrypted under stringent privacy standards and excluded from model training, medical experts advise users to treat AI as a complementary informational tool rather than a substitute for professional clinical diagnosis.

Source: TechCrunch

💻 Poolside Releases Laguna S 2.1: Open-Weight Coding Model with Advanced Self-Debugging

AI coding startup Poolside has unveiled Laguna S 2.1, its third open-weight coding model released in just three months. Rather than chasing raw parameter scale, Poolside focused on optimizing a compact architecture equipped with superior self-checking and automated debugging capabilities.

Laguna S 2.1 stands out due to its specialized training methodology: the model actively reviews its own code output, executes hypothetical test scenarios, detects and corrects logical flaws upon test failure, and discards unpromising solution paths. Benchmark evaluations demonstrate that Laguna S 2.1 achieves coding accuracy that significantly punches above its parameter weight. This milestone underscores the industry shift toward “reasoning at test-time,” offering promising applications for reliable agentic coding on local hardware.

Source: The Decoder