Archive

Today's Key Insights

  • Researchers Reverse-Engineer LLM Prompts From Output Text — Developers handling confidential prompts cannot rely on keeping target-model weights private: IIT Bombay and Adobe Research showed that output text alone can be used to reconstruct the original prompt.
  • White House Considers Adding Open Models to AI Framework — Open-model developers face uncertainty because the White House is considering adding their models to an updated framework, while WIRED reports no details on what regulation would follow.
  • Anthropic Adds Claude Cowork to Chrome, Names Legal Lead — Claude users now get a Chrome side panel, AWS’s self-hosted governance layer, and a dedicated legal lead, while Anthropic’s watermarking expansion reaches older models that some users say could expose Claude use at jobs and classes.
  • Gemini Reaches 1 Billion Users as Market Share Falls — For Google, 1 billion Gemini users now sit alongside a 1.9% Pangram share, compared with more than 50% for OpenAI and 14.9% for Anthropic.
  • Okta Targets AI Agent Token Costs With MCP Scoping — For teams running AI agents, Okta’s proposal targets the tokens consumed when models process MCP tool schemas, names, descriptions, and parameters in each call.

Top Story

Researchers Reverse-Engineer LLM Prompts From Output Text #

The Decoder AI

Researchers at IIT Bombay and Adobe Research built an inverse language model that reconstructs an original prompt from an LLM's output with near-perfect accuracy, The Decoder reports.

The method, called Previous-Token Prediction, does not require access to the target model's weights. It works backward from generated text to recover the prompt that produced it.

Why it matters: Developers handling confidential prompts cannot rely on keeping target-model weights private: IIT Bombay and Adobe Research showed that output text alone can be used to reconstruct the original prompt.

Key Takeaways

  • The researchers call the technique "Previous-Token Prediction."
  • The system is an inverse language model rather than the target LLM itself.
  • The reconstruction starts with generated output text and targets the original prompt.

Industry Updates

White House Considers Adding Open Models to AI Framework #

Wired AI

Open models may soon join an updated White House AI framework, according to sources cited by WIRED.

The possible addition comes as the White House grapples with how to regulate AI after trying not to regulate the technology.

WIRED names no specific open-model developer or application and gives no details about how the updated framework would regulate open models.

Why it matters: Open-model developers face uncertainty because the White House is considering adding their models to an updated framework, while WIRED reports no details on what regulation would follow.

Anthropic Adds Claude Cowork to Chrome, Names Legal Lead #

The Decoder AI • TechCrunch AI • AWS ML Blog

Anthropic has moved Claude Cowork into the side panel of its Chrome extension. The browser version supports skills and plugins, bringing Cowork into Chrome rather than keeping it in a separate application.

AWS has published a production reference deployment for Anthropic’s Claude apps gateway, a self-hosted governance layer that sits between Claude Code and Claude Desktop and either Amazon Bedrock or Claude Platform on AWS. AWS’s post covers the architecture, enterprise deployment patterns, cost, and implementation resources.

Anthropic has also hired legal startup founder Robert Mahari as its first “Head of Claude for Legal,” responsible for deploying and expanding Claude for legal practices. Separately, Anthropic says it will extend watermarking support to older models; some Claude users have complained that the new watermarks could catch them using Claude at their jobs or in classes.

Why it matters: Claude users now get a Chrome side panel, AWS’s self-hosted governance layer, and a dedicated legal lead, while Anthropic’s watermarking expansion reaches older models that some users say could expose Claude use at jobs and classes.

Gemini Reaches 1 Billion Users as Market Share Falls #

The Decoder AI • TechCrunch AI • Wired AI

Google says the Gemini app has reached 1 billion users, while three market-data sources report that Gemini is losing share to ChatGPT and Claude. Pangram puts Gemini's share at 1.9%, down from 12%; OpenAI holds more than 50%, and Anthropic rose from 4.3% to 14.9%. Similarweb and OpenRouter report the same direction.

At Made by Google 2026, Google announced Pixel 11, Pixel Watch 5 and Pixel Tag alongside new Gemini features. The Pixel 11 series also adds Magic Capture, Instant Night Sight and a built-in teleprompter.

Why it matters: For Google, 1 billion Gemini users now sit alongside a 1.9% Pangram share, compared with more than 50% for OpenAI and 14.9% for Anthropic.

Okta Targets AI Agent Token Costs With MCP Scoping #

AI News

Okta says identity-scoped Model Context Protocol (MCP) tool lists can reduce AI agent token costs. Each model call can include the schemas, names, descriptions, and parameters for every tool exposed by an MCP server.

Okta calls that prompt overhead the “tool tax”: tokens consumed as a model considers available tools.

Why it matters: For teams running AI agents, Okta’s proposal targets the tokens consumed when models process MCP tool schemas, names, descriptions, and parameters in each call.

Three AI Pioneers Debate Openness and Regulation at Ai4 #

TechCrunch AI

Geoffrey Hinton, Fei-Fei Li, and Andrew Ng debated regulation, open-source access, and America’s ability to compete as China advances in Asia at the Ai4 conference.

The discussion put openness and regulation at the center of a broader question facing U.S. AI leaders: how America should compete as China advances in Asia.

Why it matters: For U.S. policymakers and AI developers, the Ai4 discussion connects regulation and open-source access with America’s competition against China—three issues addressed by Hinton, Li, and Ng.