Welcome to the Agent Platform Research Briefing for Thursday, August 14th, 2026.

**DeepSeek Reverses Course — V4 Pro Launch with Up to 12x Price Increase** — DeepSeek officially released its V4 Pro 0813 model on August 13th and simultaneously announced steep API price increases taking effect August 16th UTC (August 17th Beijing time). The price war poster child is pivoting to profitability. Cache-hit input tokens for V4 Pro jump 12-fold during peak hours and 6-fold off-peak. Cache-miss input goes up 3x at peak. Output pricing rises 4.5x at peak. DeepSeek is also introducing peak and off-peak scheduling for the first time, with peak hours at 9-to-noon and 2-to-6 PM Beijing time. V4 Flash prices also increase up to 5x. This follows the V4-Flash release just two weeks ago at rock-bottom pricing of 14 cents per million tokens. Bloomberg reports the timing aligns with DeepSeek's IPO preparations — the company needs to demonstrate sustainable unit economics before going public. The era of Chinese model price disruption may be hitting a ceiling.

**MCP Security Inflection Point — Seoul Dev Summit Exposes 21,000 Exposed Servers** — The MCP Dev Summit in Seoul this week transformed from a routine industry check-in into a security reckoning. New research published alongside the OWASP MCP Top 10 framework reveals 21,000 internet-facing MCP server instances currently exposed, with 92 percent of audited production servers lacking OAuth authentication. The formal OWASP Top 10 for MCP was published this week, cataloging prompt injection, data exfiltration, and supply-chain attack vectors. The core architectural debate continues: Anthropic maintains the STDIO transport model's execution behavior is "by design" and a "secure default," placing input sanitization responsibility on developers. Security researchers counter that 687 instances have unrestricted shell tool access and the STDIO model inherently trusts local processes. The OX Security "Mother of All AI Supply Chains" report from April flagged 150 million potentially affected downstream package downloads. This is becoming the most significant security story in the agent ecosystem.

**GitHub Copilot Adds Enterprise MCP Server Allowlists** — GitHub launched centrally managed MCP server allowlists for enterprise Copilot deployments on August 6th. Enterprise owners can now use allowedMcpServers and deniedMcpServers keys in enterprise managed settings to control which MCP servers developers can run. Policies match servers by remote URL, local command, or name — with wildcard support and URL canonicalization to prevent evasion. Policies fail closed, meaning malformed configurations are blocked rather than allowed. This directly addresses the MCP security gap identified at Black Hat and follows the pattern of Azure AI Gateway and Cloudflare's Identity-Aware AI Gateway. MCP governance is maturing from "wild west" to enterprise-grade access control.

**OpenClaw 2026.8.1-beta.1 Ships** — OpenClaw released its August beta, 2026.8.1-beta.1, building on the extensive 2026.7.2 beta cycle. The new beta continues work on GPT-5.6 compatibility across OpenAI and Codex routes, improved Claude Code integration via the openclaw attach command, and broader provider support including Muse Spark 1.1, LongCat, and ClawRouter. The current stable channel remains 2026.7.1-2 with the extended-stable at 2026.6.33. The beta series now has six iterations, indicating significant internal changes before the next stable release. The openclaw attach feature — giving Claude Code temporary access to existing Gateway sessions — remains the standout feature from the 7.2 cycle.

That's the briefing for August 14th, 2026.