AI Infrastructure Matures as Open Weights and Sandboxing Take Center Stage — 2026-07-15#

Highlights#

The enterprise and infrastructure layers of AI are rapidly maturing, shifting the conversation from simple chat interfaces to robust sandboxing, automated evaluations, and embedded workflows. Meanwhile, researchers are pushing the boundaries of physical AI with test-time training for robotics, even as debates over the limits of recursive self-improvement and model supply-chain security intensify.

Top Stories#

• Perplexity Unveils SPACE Sandbox for Agents: Perplexity has shifted 100% of its Computer production traffic to a new in-house sandbox lifecycle management platform called SPACE. The system uses disposable Firecracker microVMs and Btrfs snapshots to securely pause, resume, and branch long-running agent sessions without storing credentials directly in the runtime.
• Test-Time Training Scales Robotic Context to 8,000 Timesteps: A collaboration between Stanford and NVIDIA introduces RoboTTT, leveraging a tiny neural network inside the main model to continuously compress history into weights via gradient steps on incoming sensor readings. This breakthrough enables one-shot in-context learning from human video and allows robots to recover from their own mid-episode errors.
• GPU Compute Forward Curves Launch via Prediction Markets: Tarek Mansour announced the launch of forward curves for Nvidia B200, H200, and A100 chips derived from prediction market prices. This marks a critical step toward standardizing compute as a mature commodity with its own derivative market.
• OpenAI Deploys Automated Prompt Injection Defenses: OpenAI introduced GPT-Red, an internal automated red-teamer designed to discover prompt injection vulnerabilities at scale before models see wider deployment.
• Security Researchers Sound Alarm on Open Weight Supply Chains: Security experts highlight a fundamental flaw in AI trust: unlike suspicious binaries, model weights cannot be reverse-engineered. This “mole in the model” problem is being cited as a massive vulnerability for agent security.

Articles Worth Reading#

Enterprise Agent Adoption Demands Deep Tech Integration Knowledge work lacks the immediate testability of code, making evaluations a critical bottleneck for deploying enterprise agents. According to Aaron Levie, IT leaders are discovering that successful workflow transformation requires embedding full engineers directly into business functions. Furthermore, enterprise workflows demand headless software and cross-functional agentic systems with their own discrete privileges, exposing massive security vulnerabilities in chaining mythos-level models.

Diminishing Returns Plague Recursive Self-Improvement Ramez Naam analyzes recent research on recursive self-improvement (RSI), noting that AI model advancement scales at roughly the 13th root of input intelligence. This power-law diminishing return suggests that an intelligence “explosion” or rapid software-only singularity is mathematically unlikely to materialize, as the boost from each iteration shrinks significantly.

Codebases Must Become Agent-Native Engineers are realizing that infrastructure must now capture domain knowledge previously held in human heads. By extensively utilizing documentation like CLAUDE.md, custom skills, and robust lint rules, teams can encode their project architecture so agents can immediately contribute on day one with zero additional prompting context. As workflows evolve, the ideal paradigm is shifting toward “thin prompts, thick artifacts + context, thin skills”.