AI Rundown: Anthropic-Pentagon Tensions & Qwen-3.5

Morning Brief: AI, Defense Tensions, and the Latest Model Releases

Good morning, {{ first_name | AI enthusiasts }}. Today's roundup covers a deepening feud between Anthropic and the Pentagon, new security features from OpenAI, hands-on AI design tips, and a major open-weight release from Alibaba. We also include quick notes on tools, community workflows, and other AI headlines — plus a touch of gaming news where it intersects with AI platforms and agents.

Anthropic vs. Pentagon: A Supply Chain Risk Standoff

The Pentagon may soon label Anthropic a "supply chain risk" because of the company's limits on how the military can use its Claude model. That label is normally reserved for foreign adversaries and would force U.S. defense contractors to cut ties with Anthropic.

Key facts

• Designation impact: If Anthropic is tagged a supply chain risk, defense partners would have to sever relationships, dealing a major blow to Anthropic's government business.
• Core dispute: Defense officials want permission to use AI for "all lawful purposes." Anthropic wants stricter guardrails to prevent misuse, such as spying on Americans or enabling autonomous weapons.
• Current deployments: Claude is the only AI reported on Pentagon classified systems. It was also reportedly used via Palantir in operations such as the Nicolás Maduro capture in January.

Why it matters: This standoff highlights a broader debate in the AI era — who controls frontline models in military operations: the labs that build them or the governments that deploy them?

Experts warn that unchecked AI use in warfare risks serious ethical and safety problems. The outcome of this dispute could set precedents for future national-security AI governance.

You.com Predicts an AI Winter

You.com co-founders Richard Socher and Bryan McCann released 35 predictions for 2026. A few notable forecasts:

• LLM returns diminish as capital shifts back to fundamental research.
• "Reward engineering" becomes a specialized job beyond prompting.
• Traditional coding largely replaced by code; humans manage systems instead of writing low-level routines.

OpenAI Adds Lockdown Mode to ChatGPT

OpenAI launched a new optional "Lockdown Mode" in ChatGPT to protect security-conscious users from threats like prompt injection. The company also introduced "Elevated Risk" labels across several products.

What Lockdown Mode does

• Deterministically disables certain tools and capabilities that attackers could exploit.
• Limits web browsing to cached content to prevent live network requests leaving OpenAI's environment.
• Allows ChatGPT workspace admins to whitelist specific apps/actions when lockdown is active.

Why it matters: As chatbots evolve into agents that browse, connect to apps, and act on behalf of users, deterministic hard blocks may be necessary to manage security risks.

AI Training: Turn a YouTube Thumbnail into Five Social Posts

Quick guide: build a YouTube thumbnail in Canva, then use its AI resize feature to generate optimized versions for every social platform.

1. In Canva (Pro account required), click Create and choose "YouTube thumbnail."
2. Select a template. Update the title text and colors.
3. Replace any placeholder cutouts by uploading a selfie and using AI background removal.
4. Click Resize and select desired social placements. Canva AI will adapt layouts for each size.

Pro tip: Generate initial concepts with Nano Banana Pro, then refine in Canva for consistent branding across platforms.

Sponsored: Speechmatics — Fast, Trusted Voice Agents

Speechmatics advertises sub-250ms partials and 250–300ms finals for real-time voice agents. Supports 55+ languages and flexible deployment (cloud, on-prem, hybrid, on-device).

Alibaba's Qwen-3.5: Open-Weight, Sparse MoE, Near-Frontier Performance

Alibaba released Qwen3.5-397B-A17B, an open-weight vision-language model that uses a hybrid sparse Mixture-of-Experts (MoE) architecture. It activates 17B parameters per query from a 397B pool.

Highlights

• Performance: Competes with models like GPT-5.2 and Gemini 3 Pro in agentic search, document recognition, and instruction following.
• Efficiency: Alibaba claims 60% lower cost and 8x better large-workload processing than its predecessor.
• Use cases: Designed for continuous multimodal reasoning for agents and large-scale inference workloads.

Why it matters: Open-weight models that balance cost, latency, and capability can shift the competitive landscape. Labs that deliver powerful, efficient models at lower costs will press an advantage.

Trending AI Tools

• HeyGen — Create videos from ideas without filming. Promo code: TRDAI20 for a discount.
• Manus — AI agents available via Telegram with long-term memory.
• Seed 2.0 — ByteDance's open-source models designed for general-purpose agents.

Other AI Headlines (Short Takes)

• Meta patented a system to simulate a user's responses using AI trained on their interaction data, including options for long breaks or posthumous simulation.
• India hosted an AI Impact Summit with leaders from OpenAI, Google, and Anthropic. Anthropic announced a Bengaluru office; India is now the second-largest market for ChatGPT and Claude.
• Ireland's Data Protection Commission probes xAI's Grok over image-generation concerns; similar scrutiny continues in the UK and EU.
• SpaceX/xAI may compete for a Pentagon contest to build voice-controlled autonomous drone swarms.
• ElevenLabs launched "ElevenLabs for Government" to help public agencies deploy secure, multilingual voice and chat AI.

Community AI Workflow

Reader spotlight: Lisa G. from Australia volunteers at a wildlife shelter and built a rescue-management app using Base44's web app creator. The app tracks rescue calls, volunteer tasks, and mapping, replacing inefficient WhatsApp threads.

Want to share your AI workflow? Tell us here.

Gaming News and AI

Gaming news increasingly intersects with AI — from NPC behavior driven by large models to voice agents and content moderation systems. Expect more games to use AI agents for dynamic storytelling, procedural content generation, and adaptive difficulty.

Wrap-Up

This moment in AI blends ethical debate, rapid model advances, and growing security needs. The Anthropic-Pentagon dispute, OpenAI's Lockdown Mode, and Alibaba's Qwen-3.5 show how companies, governments, and developers are racing to define where responsibility and capability meet.

See you soon, Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown