This Week in AI (7/28 - 8/2)

TL;DR

• Anthropic blocks OpenAI from accessing Claude, citing terms violations over benchmarking.

• GPT-5 is on the horizon: Multimodal, longer context, and fewer hallucinations—expect changes to your AI workflows and budgets.

• EU AI Act’s new GPAI rules now apply; documentation and compliance are essential for anyone operating in Europe.

• Skyflow launches MCP security layer: More focus on securing agent–tool integrations as adoption grows.

Anthropic Blocks OpenAI’s Access to Claude

On August 1, Anthropic revoked OpenAI’s API keys for Claude—specifically targeting “Claude Code” calls—after catching OpenAI engineers benchmarking Claude in advance of the upcoming GPT-5 launch (WIRED, WinBuzzer, THE DECODER, WebProNews, BleepingComputer).

• Anthropic’s rationale: Protect their models from competitor benchmarking; enforce terms of service.

• OpenAI’s response: Says benchmarking rivals is industry standard for safety and performance.

  
  

Why it matters: Claude Code is clearly a core moat for Anthropic, and this move reflects concerns about losing that advantage. If your team benchmarks against Claude, expect slower cycles or tighter licensing terms. Consider open-source or less restrictive models as alternatives.

GPT-5 Release Imminent

Insiders—and Sam Altman himself—suggest GPT-5 could land as soon as August 2025 (The Times of India, The Verge).

• Unified multimodal reasoning: Handles text, code, images, and more.

• Longer context windows: Rumored 200k+ token limits.

• Reduced hallucinations: Fewer made-up answers; more reliability for enterprise use.

• Initial rollout: Expected behind ChatGPT Plus and Enterprise paywalls.

  
  

Why it matters: GPT-5 hype is high—arguably higher than any model can deliver on at launch. Still, OpenAI’s aiming to reclaim the lead. Start reviewing your stack and prototyping now, but be realistic: not every promised feature will make an immediate impact, and migration may take time.

EU AI Act’s GPAI Rules Now in Force

As of August 2, new obligations under the EU AI Act’s “General-Purpose AI” (GPAI) rules are enforceable (IT Pro, Digital Strategy, Mayer Brown):

• Model documentation: Required disclosure of data sources, methods, and use cases.

• Transparency reports: Ongoing publication of model performance and risks.

• Safety-by-design checks: Mandated bias and robustness audits.

  
  

Why it matters: While the EU is leading on AI regulation and transparency, some worry this could drive companies out of the market as the US and China remain more hands-off. If you handle EU user data, compliance is mandatory. Some firms may opt to withdraw, or wait and see how enforcement plays out.

Skyflow Unveils MCP Data Protection Layer

On August 1, Skyflow announced its MCP Data Protection Layer, a new security solution for agentic workflows using the Model Context Protocol. The product adds encryption, fine-grained access controls, and audit logging to MCP-enabled connections, targeting SaaS vendors and enterprise deployments (Silicon Canals).

Why it matters: As organizations rush to adopt MCP, security concerns have increased. A dedicated layer for protection is overdue. Many in the field will be watching to see if native security features become standard in MCP as usage expands.

Quick Takeaways

• Benchmarking rules are shifting: Anthropic–OpenAI tensions will ripple through how teams compare models.

• GPT-5 could set new norms: But expect some gap between launch promises and practical gains.

• Compliance is non-optional: EU rules are now real; review your processes if you serve Europe.

• Security is in focus: MCP adoption brings new risks and more attention to safe integration.

  
  

See Ya!

Another week, another round of headlines. Competition, regulation, and security are all moving targets. Worth keeping an eye on what actually sticks in the coming months.

✌️ Steven