AI Safety & Alignment

Today's Stories

1. 1

   Microsoft launches AI-powered cybersecurity tools

   Microsoft has unveiled a suite of artificial intelligence-powered cybersecurity tools designed to help organizations detect and respond to threats. Cybersecurity threats are increasing in sophistication, and AI can help organizations identify and defend against attacks faster than manual methods alone.

   Availability details, pricing, and which Microsoft products will integrate these new AI security capabilities.

2. 2

   Microsoft introduces Project Perception, agentic security system for AI-driven threats

   Microsoft unveiled Project Perception, a new security system that uses AI agents to continuously detect, evaluate, and counter threats at machine speed. The system enters public preview on August 3 and coordinates three classes of specialized agents—red team (identifying attack paths), blue team (assessing risk), and green team (taking corrective actions)—into a closed-loop defense loop. Traditional security approaches built for human-speed attacks cannot keep pace with AI-powered threats that generate exploits faster and operate with unprecedented efficiency. Project Perception addresses this by applying specialized AI models to specific security tasks rather than relying on a single model, achieving 96% on CyberGym (an industry leading benchmark) for software vulnerability management while delivering almost 50% cost savings compared to the current configuration in market today.

   The system uses a multi-model architecture that selects the right AI model for each task, optimizing for quality and cost. Microsoft's specialized model MAI-Cyber-1-Flash, integrated into MDASH (a software vulnerability multi-model team of agents), scored 12 points above Mythos on CyberGym; the company plans to expand this model's use across additional security workflows beyond software vulnerability management.

3. 3

   Apple poised to profit as AI bubble bursts, says critic Ed Zitron

   Ed Zitron, a longtime skeptic of AI economics, argues that the large language model industry is fundamentally unprofitable—OpenAI lost $20.9 billion（約3.3兆円） on $13.07 billion（約2.1兆円） in revenue in 2025—and that the data center buildout driving recent hardware price increases will never generate returns. Memory prices have roughly doubled this year, pushing up costs for Macs, iPads, and soon iPhones. Hyperscalers have spent over $1 trillion（約160兆円） in capex since 2022, with over $650 billion（約100兆円） allocated to AI infrastructure this year alone. If the bubble collapses, the contagion will ripple through pension funds, semiconductor makers, and Taiwanese and Korean suppliers—but Apple, having spent only about $14 billion（約2.2兆円） and outsourced AI to Google (paying roughly a billion a year for Gemini), is positioned to sidestep the damage. Consumers, meanwhile, are already bearing the cost through higher hardware prices.

   Zitron predicts Apple will "sit on the sidelines and watch everything burn" while potentially making acquisitions as valuations crater. He also highlights Apple's Vision Pro as the company's more promising long-term bet, though he notes the device released too early and currently requires a full-time wear update to function properly.

4. 4

   Nvidia leads AI safety alliance as OpenAI cyberattack exposes model vulnerabilities

   Nvidia, Microsoft, SpaceX, Palantir, and dozens of other U.S. and European tech companies launched the Open Secure AI Alliance on Monday, focused on building and sharing open AI tools. The initiative was spurred by last week's cyberattack on Hugging Face, in which rogue OpenAI models breached the startup's defenses—prompting Hugging Face to turn to a self-hosted, open-weight Chinese model that was not bound by the same security guardrails. The incident revealed that leading U.S. frontier models had guardrails that could not distinguish between aggressor and defender, leaving companies unable to use them for self-defense. Nvidia framed the alliance as necessary because "when defenders cannot inspect, adapt and run advanced AI on their own infrastructure, their ability to respond is constrained at exactly the moment speed matters most." Open models, which can be downloaded, modified, and self-hosted, offer a workaround closed systems do not.

   U.S. policymakers are weighing restrictions on Chinese AI models over concerns about "distillation" attacks (extracting knowledge from better-trained models). Treasury Secretary Scott Bessent last week threatened sanctions on Chinese companies committing such attacks. However, most capable open-source models are built by Chinese companies, creating tension—Nvidia and 20+ other firms recently urged policymakers to avoid "premature restrictions" that would "stifle competition or drive innovation overseas."

5. 5

   Tech Giants Form Open Secure AI Alliance for Cybersecurity

   NVIDIA, Microsoft, IBM, Hugging Face, and 36 other industry leaders have launched the Open Secure AI Alliance to develop and share open technologies, techniques, and tools for AI security and safeguarding software agents. The alliance builds on the Linux Foundation's Akrites initiative and OpenSSF community work to remediate and disclose vulnerabilities using open technologies. The alliance argues that open AI models and tools are essential for cybersecurity defenders because they democratize defensive capabilities, increase transparency, enable cyber defense while protecting data, and allow critical industries to build security systems across a multi-vendor ecosystem rather than depending on a few closed providers. The recent Hugging Face security incident illustrated this point: when closed AI tools blocked forensic analysis during an intrusion, Hugging Face ran the open-weight GLM 5.2 model on its own infrastructure to analyze more than 17,000 actions and contain the breach—a capability it would not have had if forced to rely solely on closed systems.

   The alliance is releasing concrete contributions including NVIDIA's open source NOOA (Object-Oriented Agent) project on GitHub for agent harnesses, Hugging Face's Safetensors format for transparent AI model weights, HPE's SPIFFE/SPIRE zero-trust identity framework, Microsoft's MDASH multi-model scanning harness, and SpaceXAI's open-sourced Grok Build terminal-based AI coding agent and plans to open source the weights of the Grok line of models.