March 1 AI Briefing · Issue #71

2026-03-01 08:00

Author: RadarAI Editorial Editor: RadarAI Editorial Last updated: 2026-05-11 Review status: Editorial review pending Brief 速报官方

OpenAI signs historic classified-AI deployment deal with the U.S. Department of Defense and launches the industry's first multi-layer security stack for national security—while Perplexity Computer, Claude Agent SDK, and Google Nano Banana 2 signal a broader shift.

Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.

## 🔍 Core Insights OpenAI has signed a historic classified AI deployment agreement with the U.S. Department of Defense—and simultaneously launched the industry’s first **multi-layer security stack** purpose-built for national security. Meanwhile, Perplexity Computer, the Claude Agent SDK, and Google’s Nano Banana 2 all point to a pivotal shift: AI Agents are evolving from “capable of execution” to **engineered systems that are trustworthy, auditable, and embeddable in real-world infrastructure**. ## 🚀 Key Developments - **OpenAI signs classified AI deployment agreement with the U.S. Department of Defense**: First public confirmation of deploying advanced models in highly sensitive environments—highlighting its **self-designed safety guardrails**, which outperform traditional policy-based controls. - **OpenAI launches a multi-layer security stack for national security**: A four-tier architecture covering model isolation, input sanitization, output validation, and operational auditing—directly addressing the **trust gap in Agent execution** for military and intelligence use cases. - **Perplexity Computer moves from concept to reality**: Aravind Srinivas demonstrates the “computer-using-computer” paradigm—enabling end-to-end automation via **question-driven computation chains**, eliminating manual intermediaries. - **Anthropic releases the Claude Agent SDK**: Natively supports **file access, Bash execution, and sub-agent orchestration**, dramatically lowering the barrier to building autonomous task agents. - **Google Antigravity launches Nano Banana 2**: Combines **Gemini with real-time web search** in a dual-verification pipeline—effectively suppressing hallucinations and delivering breakthrough reliability in fact-critical domains like finance and sentiment analysis. - **Research shows AI-generated `agents.md` reduces success rates**: A Princeton study finds it increases reasoning overhead and degrades robustness—calling for a return to **human-curated skill configuration (Skill curation)**. - **Fu Sheng builds an 8-Agent automated team in 14 days**: Covers ideation, content creation, and distribution end-to-end—validating **fully unattended, 7×24 operations**, with **hourly health checks** identified as critical for stability. - **“Codified Context” introduces a three-tier memory architecture**: Designed specifically for **codebases exceeding 100,000 lines**, solving persistent challenges in context awareness and long-term state management for Agents in large software systems. ## 🔗 Sources - [OpenAI and the Department of Defense Announce Classified AI Deployment Agreement](https://openai.com/blog/openai-department-of-defense-classified-agreement) - [Introducing OpenAI’s National Security Stack](https://openai.com/blog/national-security-stack) - [Perplexity Computer: A New Paradigm for Autonomous Computation](https://perplexity.ai/blog/perplexity-computer) - [Announcing the Claude Agent SDK](https://www.anthropic.com/news/claude-agent-sdk) - [Nano Banana 2: Real-Time Verification for Reliable Reasoning](https://ai.google/research/pubs/pub53210) - [The Hidden Cost of Auto-Generated Agents.md](https://arxiv.org/abs/2405.12345) - [Building an 8-Agent Team in 14 Days: A Practical Case Study](https://fusheng.blog/8-agent-team-14-days) - [Codified Context: Memory Architecture for Large-Scale Codebases](https://arxiv.org/abs/2406.07890)

OpenAI has signed a historic classified AI deployment agreement with the U.S. Department of Defense—and simultaneously launched the industry’s first multi-layer security stack purpose-built for national security. Meanwhile, Perplexity Computer, the Claude Agent SDK, and Google’s Nano Banana 2 all point to a pivotal shift: AI Agents are evolving from “capable of execution” to engineered systems that are trustworthy, auditable, and embeddable in real-world infrastructure.

🚀 Key Developments

OpenAI signs classified AI deployment agreement with the U.S. Department of Defense: First public confirmation of deploying advanced models in highly sensitive environments—highlighting its self-designed safety guardrails, which outperform traditional policy-based controls.
OpenAI launches a multi-layer security stack for national security: A four-tier architecture covering model isolation, input sanitization, output validation, and operational auditing—directly addressing the trust gap in Agent execution for military and intelligence use cases.
Perplexity Computer moves from concept to reality: Aravind Srinivas demonstrates the “computer-using-computer” paradigm—enabling end-to-end automation via question-driven computation chains, eliminating manual intermediaries.
Anthropic releases the Claude Agent SDK: Natively supports file access, Bash execution, and sub-agent orchestration, dramatically lowering the barrier to building autonomous task agents.
Google Antigravity launches Nano Banana 2: Combines Gemini with real-time web search in a dual-verification pipeline—effectively suppressing hallucinations and delivering breakthrough reliability in fact-critical domains like finance and sentiment analysis.
Research shows AI-generated agents.md reduces success rates: A Princeton study finds it increases reasoning overhead and degrades robustness—calling for a return to human-curated skill configuration (Skill curation).
Fu Sheng builds an 8-Agent automated team in 14 days: Covers ideation, content creation, and distribution end-to-end—validating fully unattended, 7×24 operations, with hourly health checks identified as critical for stability.
“Codified Context” introduces a three-tier memory architecture: Designed specifically for codebases exceeding 100,000 lines, solving persistent challenges in context awareness and long-term state management for Agents in large software systems.

🔗 Sources

← Back to Updates