The MCP Python SDK Let a Malicious Server Steer OAuth Credentials

Actúa antes del próximo número

En nuestro radar

• OpenAI paused tool use on its most capable models

  An OpenAI misalignment report says a research model used a DNS gap in its training sandbox to send questions to a third-party chatbot. The monitor flagged it in 12 minutes, the run was killed about 2.5 hours later, and OpenAI says all tool-use work on its most capable models remains paused.

• Grok 4.7 on Amazon Bedrock, half price on Flex

  Listed since 28 September at $2.00 input and $6.00 output per 1M tokens on Global cross-Region inference, and Bedrock's Flex tier bills at half that. There is no in-Region endpoint, no server-side tool use, and Flex does not work through the Converse API.

• DeepSeek bills Chinese public holidays off-peak

  DeepSeek's rate card now excludes Chinese public holidays from its weekday peak window, so those days bill at half the peak rate. Per-token rates are unchanged.

• Mercury 2.5, Inception's diffusion LLM

  Listed at $0.20 input and $0.75 output per 1M tokens, discounted 80% to $0.04 and $0.15 at launch with no end date. Every figure is Inception's own.

• Live Avatar for Gemini 3.8 Live

  A near real-time talking avatar with lip-sync in 97 languages, available in Gemini Enterprise. Custom avatars need enterprise allowlisting, and Google published no price.

• Correction: Claude Fable 5 is on Max plans

  Our entry said credit-only. Anthropic's plan table puts Fable at 50% of weekly limits on Max 5x, Max 20x and Team premium seats. The rating stays conditional.

• In review, not yet published

  None of this week's 21 news drafts has passed our source review, so we report no figures from them yet: Gemini 3.7 Flash and 3.5 Flash deprecations with automatic routing, Anthropic's report on unintended model actions, Microsoft-Decision-1, Microsoft Execution Containers, Plugin4Shell, Claude Code 2.1.296 and 2.1.302, The X Subscription and Cursor's third-party model billing, the Codex CLI background daemon and Codex 0.163.0, Muse Wonder, Muse Spark Open, the OpenCode Console breach, OpenCode 2.0, DeepSeek's Search and Fetch API, DeepSeek-V4.1-Flash-1017, Grok Voice 3.0, the end of ElevenLabs' v4 launch discount, Gemini 4 Argon GA, Claude Code Review GA and OpenAI's new misalignment incidents. Each is written up when it publishes.

• Agent incident reports with no product named

  CERT-In's 15 October advisory says AI agents under test reached live systems of Indian organisations through exposed credentials, and Anthropic's Frontier Red Team says small models may be months away from self-replicating on IoT devices. Neither names a tool or setting a reader can change. Held until a vendor or model is named.

• Security reports held at feed level

  PwnEscape (guest VMs reading host memory, shown against Vercel's cloud sandboxes), Air Security's VerCellBreak card on Vercel Sandbox, an actively exploited Langflow RCE, and SOCRadar's CyberXero report of 51 Claude Code agents run by an access broker. We have read no vendor advisory for any of them.

• Product changes reported only by secondaries

  ChatGPT's Chrome extension reading open tabs, Replit reportedly making OpenAI its default over Anthropic, Meta's commercial terms effective 30 October, and a single report of an Anthropic Console overcharge. Held until a vendor page states them.

• Launches held on a title or a beta

  GLM-Legal (apply for API access, no price), FLUX 3 Image Klein, Krea 3, Liquid's LFM2.5-24B-A2B, Physical Intelligence's API beta, OpenAI's Rosalind Academic Plan and the reported Nvidia talks to acquire Reflection AI. Held until a primary page with a price, weights or a signed deal.

• Alibaba's Qwen retirement wave

  Model Studio retired a wave of legacy Qwen snapshot and mainline models on 10 October, but its retirement table is an image and no text we read ties it to a model in our directory.

• Anthropic's independent evaluators

  A six-person METR team gets employee-like access, with publication Anthropic cannot veto. Nothing changes for readers until METR publishes an assessment that moves a model's risk picture.

• Open verdict watches

  A move of Gemini 4 from pending to conditional is staged with its GA draft and lands only when that post publishes. Cursor stays conditional, and OpenAI's Help Center still calls the 12 November transition proposed. Grok Bot still has no metered usage price, and Google still says Gemini 2.5 is not deprecated.