AI News Roundup: August 26-27, 2026 — Claude 4.7 Opus T-Minus 9 Days, OWASP Agentic Top 10 v2.0 Ships, and Agent Platform Wars Heat Up
Models & Releases
Claude 4.7 Opus ("Mythos 4.7") Officially Announced for September 5
With just 9 days until launch, the Claude 4.7 Opus countdown intensifies. Anthropic confirmed via internal communications that "Mythos 4.7" will debut on September 5, aligned with the company's "Agent Week" conference. The model ships with a 2M token context window — doubling the current 1M ceiling — and native video understanding, eliminating the need for separate vision models in video analysis workflows.
For agent builders, the implications are significant: a 2M context window allows entire codebases or multi-thousand-page document corpora to fit in a single context, removing the chunking overhead that currently limits agent reasoning depth. Native video understanding opens entirely new use cases in automated video content review, video-based research, and multimedia QA workflows.
Anthropic's internal documentation suggests Agent Week will also debut new tooling for the Agent Skill Protocol ecosystem, though specific announcements remain under wraps.
DeepSeek V4.5 Flash Surpasses GPT-5 on MMLU-Pro
DeepSeek published benchmark scores for V4.5 Flash (Apache 2.0 licensed): 89.2 on MMLU-Pro, surpassing GPT-5 at 87.5 while maintaining a 3x inference speed advantage over GPT-5.6 Sol standard mode. The model remains free via DeepSeek's API for up to 1M tokens/day. This continues DeepSeek's relentless trajectory of delivering frontier-model performance at a fraction of the cost, pressuring US labs on price-performance metrics.
Agent Infrastructure
OWASP Agentic AI Security Top 10 v2.0 Released
The OWASP Agentic AI Security Project's v2.0 release is now live, adding three critical attack categories absent in v1.0:
- A06: Prompt Injection via Tool Output — Malicious content injected through tool results, bypassing traditional input filtering
- A07: Agent Privilege Escalation — Agents gaining unauthorized access to sensitive tools through chain-of-thought manipulation
- A08: Memory Poisoning — Adversarial manipulation of long-term agent memory to corrupt future decisions
The update reflects the rapid maturation of agent systems from research prototypes to production deployments where vulnerabilities are increasingly weaponized. Tencent's AI-Infra-Guard, now at 5,822 stars, specifically addresses these categories through an automated scanning pipeline.
Cursor Agent SDK (cursor/plugins) Crosses 5K Stars
Cursor's open-source plugin SDK for its code hosting platform reached 5,074 stars (+1,832 this week) as Cursor crossed 1M hosted repositories. The SDK defines the plugin specification for Cursor's expanding agent platform, enabling PR-triggered AI workflows including code review, automated testing, and security scanning.
The rapid adoption signals a strategic shift: Cursor is positioning itself as an AI-native development platform, not just a coding assistant. The SDK's event-driven architecture (triggered by PR, push, and issue events) creates a composable agent pipeline that competitors like GitHub are now scrambling to match.
Databricks Agent Orchestration GA at $188B Valuation
Databricks reached its $188B valuation milestone following the GA release of its Agent Orchestration platform, built on Apache OpenLineage. The platform brings enterprise-grade governance — Unity Catalog access control, MLflow experiment tracking, and per-agent cost attribution — to production AI agent deployments. Early customers include Goldman Sachs, Merck, and UPS.
The OpenLineage foundation ensures interoperability: agents can be observed and managed through any OpenLineage-compatible tooling, not just Databricks. This open approach contrasts with more closed platforms and may become a differentiator in the enterprise agent infrastructure market.
Memory-Interop MCP: Cross-Framework Agent Memory
The memory-interop MCP server reached 2,100 stars, offering a unified memory layer across LangGraph, CrewAI, and AutoGen agents. By exposing memory as MCP tools, the project enables agents from different frameworks to share and search memory stores (Qdrant, Pinecone, Weaviate, and 9 more backends).
This addresses a critical pain point: as organizations deploy agents across multiple frameworks, memory fragmentation becomes a scalability bottleneck. Memory-interop's MCP-native approach provides a protocol-level solution rather than another framework-specific abstraction.
Security & Governance
China's AI Agent Export Controls Enter Public Comment Period
China's Ministry of Commerce opened public comment on draft regulations classifying AI agent systems with >10B parameters as dual-use technology, open until October 1. The rules would cover agent frameworks, training data, and fine-tuned models — China's first AI-agent-specific export control regime.
MCP 2.2 Draft: Streaming Results and Cross-Server Chaining
The MCP working group published the MCP 2.2 draft, introducing streaming tool results (reducing perceived latency for long-running operations) and cross-server chaining (enabling one MCP server to invoke tools from another). These features are critical for building complex multi-agent systems where agents collaborate through shared tool interfaces.
GitHub Trending — AI Agents (Week of Aug 22-27)
| Project | Stars | Language | Category | Weekly Gain |
|---|---|---|---|---|
| OpenViking | 33.1K | Python/Rust | Agent Memory & RAG | +4.0K |
| cursor/plugins | 5.1K | TypeScript | Agent Platform SDK | +1.8K |
| omlx | 20.6K | Python | Apple Silicon Inference | +1.7K |
| Tencent/AI-Infra-Guard | 5.8K | Go/Python | AI Security | +1.2K |
| Apache Maka | 3.1K | TypeScript | Agent Workspace | +1.3K |
| akitaonrails/ai-memory | 4.5K | Rust | Agent Memory | +2.5K |
| nodeterm | 1.2K | TypeScript | Agent Terminal | +529 |
| claude-plugins-community | 1.5K | Python | Agent Skills | +877 |
Broader Trends
The week's dominant narrative is agent infrastructure consolidation. Databricks achieving GA at $188B, Cursor launching a full agent SDK, and MCP 2.2 adding cross-server chaining all signal that agent infrastructure is moving from fragmented open-source experiments into a period of platform competition with clear enterprise winners.
The OWASP v2.0 release and China's export controls add a governance layer to this consolidation — as agents move into production, the questions of security and regulation are catching up to the technology. September 5's Claude 4.7 Opus launch will be the defining inflection point of this cycle: a 2M context window and native video understanding could redraw the competitive landscape for autonomous agent capabilities.