October 07, 2026

7 research entries this day.

2026-10-07Safety & Risk

Meta, Sierra, Walmart and Stripe Propose a 'Personal Agent Protocol' So Businesses Can Tell Which Bot Is Acting for Whom

On October 6, Sierra and Meta announced Personal Agent Protocol, an open standard being developed with Genesys, Instinct, Rocket, Shopify, Stripe and Walmart. It gives a personal agent like Meta's Muse an OAuth-based session with a business, with the customer choosing read-only or write access and the company setting what the agent may do. The v0.1 spec is due later this month; OpenAI and Anthropic have not signed on.

2026-10-07Enterprise Adoption

SAP Takes Joule Out of Its Apps: Joule Work Starts Customer Rollout, Already Live for 110,000 SAP Employees

At SAP Connect on October 6, SAP began rolling out Joule Work, a full-screen AI interface that answers questions across SAP and third-party systems, and said its Autonomous Enterprise architecture becomes generally available this month. SAP says the tool is live for 110,000 of its own staff; Nestlé, Novartis, PwC and Morgan Foods described early uses. Pricing is consumption-based.

2026-10-07Research Trends

Google's EmbeddingGemma 2 Puts Text, Code, Images, Audio and Video in One On-Device Embedding Space -- From 270M Parameters

Google DeepMind released EmbeddingGemma 2 on October 6: an Apache 2.0 open embedding model built on Gemma 4 whose modular encoders let an app load only the modalities it needs, from a 270M text-and-code core to 740M for everything. Code retrieval jumps almost 10 points on MTEB Code, and Matryoshka truncation cuts vector storage up to six times.

2026-10-07Research Radar

Research Radar: Reflection's Beam and Mistral Large 4 Arrive a Day Apart -- Both Built on Asynchronous RL at Scale, Neither With Weights Yet

Reflection announced Beam, a 501B-parameter (23B active) text model, on October 5; Mistral opened a preview of Mistral Large 4 on October 6. Both describe the same recipe -- reinforcement learning with rollout generation and training decoupled -- and both promise open weights later this month. Until then, every benchmark is the vendor's own.

2026-10-07Open Source & GitHub

GitHub Trending: REA Gives Coding Agents a Reverse-Engineering Toolkit -- 'See a Feature You Like' Down to the Binary

morluto/rea drew the most new stars of any repository on GitHub's daily trending page on the morning of October 7 (IST): 2,956. It is an MCP server that lets Claude Code, Codex, Cursor and others inspect native binaries, Electron apps, .NET assemblies and websites through Ghidra, Hopper or IDA. Skill packs from Matt Pocock and an 'ADHD-friendly output' skill also climbed.

2026-10-07Agent Frameworks & Harnesses

Framework Watch: Claude Code 2.1.292 Closes a UNC-Path Prompt Bypass; Copilot Studio Hooks Arrive -- and Fail Open

Claude Code 2.1.292 (October 6) adds a per-subagent effort setting and fixes a set of permission edges, including PreToolUse hook approvals and auto mode skipping the prompt for reads from network paths. Microsoft's new Copilot Studio hooks, reported October 6, give low-code agents lifecycle events -- but only the pre-tool event can block, and a failed hook lets the agent carry on.

2026-10-07Trend Watch

Trend Watch: Anthropic Opens Tiered Cyber Access the Same Day Mistral Sells Open Weights on Closed Models' Refusals

On October 6, Anthropic merged Project Glasswing into a three-tier Cyber Verification Program that unlocks progressively fewer blocks for vetted defenders, red teams and critical-infrastructure testers. The same day, Mistral launched Large 4 by pointing out that Claude Opus 5.5 and GPT-6 Astra score near zero on a vulnerability-reproduction test because they refuse. Two answers to one question: who gets cyber capability, and who decides.