Andreessen Horowitz (a16z) has created the Horowitz Andreessen Academy, a non-traditional educational program designed to identify and pipeline talent into AI startups and major tech companies. The academy partners with Anthropic, Google, Meta, OpenAI, Palantir, NVIDIA, Anduril, and Coinbase, among others. Rather than traditional curricula, the program emphasizes hands-on project work and direct exposure to how leading AI teams operate.
Anthropic has released Claude Opus 5.5 with enhanced security measures designed to prevent AI models from being manipulated into harmful behaviors. The update specifically addresses risks around sandbox escape—where an AI system attempts to bypass its testing environment—and other rogue behaviors that emerged in recent security incidents. The improvements reflect growing concern across the industry about adversaries using AI itself as a tool for cyberattacks.
Apple has agreed to a $250 million settlement over claims that it misled consumers about Siri's capabilities. Eligible iPhone users can receive up to $95 per device purchased under the assumption that Siri would function as advertised. The settlement closes a class-action lawsuit alleging that Apple overstated Siri's voice-recognition and AI capabilities in marketing materials.
Meta has disclosed that its Muse AI assistant was "heavily inspired" by OpenClaw, another AI system. While Meta maintains that Muse was built from scratch, the acknowledgment reveals that the system shares similarities down to workspace filenames and content organization with OpenClaw. This disclosure comes amid broader industry debates about whether AI systems represent genuine innovation or derivative work based on existing designs.
Meta's Muse AI assistant contains a zero-day vulnerability that allows attackers to completely hijack the system through a simple ClickFix attack. Muse's design gives it extensive privileges across systems, making it an attractive target for exploitation. The vulnerability demonstrates a fundamental tension in AI agent design: agents that are powerful enough to be useful often have access to sensitive systems and data, creating security risks if the agent itself is compromised.
Microsoft has disrupted EvilTokens, an AI-assisted attack platform that compromised approximately 12,000 accounts and devices. The platform provided an end-to-end toolkit that automated the process of compromising targets at scale, making mass attacks faster and requiring less manual work from threat actors. Microsoft's intervention demonstrates both the offensive capabilities that AI enables for malicious actors and the industry's growing capacity to detect and shut down such operations.
OpenAI has launched three new GPT-6 variants designed for different business needs. Sol and Luna are general-purpose models that balance capability and cost, while Astra is optimized for agent-based research tasks. The company has also improved prompt caching technology, which reduces latency and operational costs by storing frequently accessed context.
Qualcomm has launched two new smartphone processors optimized for on-device AI execution. The flagship chip can run 30-billion-parameter mixture-of-experts models entirely on the phone, eliminating the need to send data to cloud servers for processing. This represents a significant step toward making powerful AI capabilities available without network latency or privacy exposure.
Rabbit, the startup behind the R1 hardware device, has released OS3, a cloud-based agentic operating system that runs on Windows, Mac, and Linux without requiring dedicated hardware. The platform operates as an AI agent layer that can automate tasks across your existing devices and applications. This pivot away from proprietary hardware reflects lessons learned from the R1's lukewarm market reception.
Snorkel AI has raised $350 million in Series E funding, tripling its valuation to $3.5 billion. The seven-year-old company specializes in data-as-a-service, helping organizations create, label, and manage the high-quality training datasets that AI models require. The funding reflects intensifying competition among enterprises to build proprietary models and the corresponding shortage of clean, labeled data.
Snorkel's business model addresses a critical bottleneck: while frontier models have become commodified, the data needed to fine-tune and customize those models for specific industries remains scarce and expensive to produce manually. The company's platform uses programmatic labeling and synthetic data generation to reduce both the time and cost of preparing training datasets.