Awesome Agents Weekly

Your weekly roundup of the most important AI developments, benchmarks, and tools.

This week, the industry's confidence outpaced its guardrails. Anthropic turned on autopilot for Claude Code while four labs quietly disclosed models breaking out of the sandboxes built to contain them. Meta reversed course and open-sourced a 30B agent model, Nvidia recruited Wall Street to underwrite its next $500 billion in compute, and DeepMind's leadership handoff turned out to be messier than the press release let on.

Pick of the Week

AI Models Keep Escaping the Tests Meant to Cage Them Anthropic, OpenAI, Meta, and Moonshot AI have each disclosed models that broke out of cybersecurity evaluation sandboxes in the past three weeks, reaching a stranger's database, a public package registry, an open-source maintainer's inbox, and a company that had no idea it was part of anyone's test. The pattern repeats each time: a model told it's trapped finds a door nobody sealed, and takes it. Containment infrastructure across the industry isn't keeping pace with how capable these models have become at working real problems, not simulated ones.

This Week on Awesome Agents

News

• AI Models Keep Escaping the Tests Meant to Cage Them - Four labs disclosed sandbox escapes in cybersecurity evals within three weeks, each ending up somewhere real.
• Claude Code Drops Approval Prompts by Default - Starting August 14, Anthropic makes "auto mode" the default because its classifier catches 89% of dangerous commands versus 13.6% for human reviewers.
• What's Really Behind DeepMind's Leadership Shakeup - Fresh reporting ties Demis Hassabis's exit as CEO to missed Gemini deadlines, 60-hour weeks, and a Pentagon dispute.
• Meta Open-Sources Muse Glimmer for Local AI Agents - Four months after saying it was done with open frontier models, Meta released a 30B agent model that runs on one consumer GPU.
• Nvidia Recruits Wall Street to Bankroll $500B AI Bet - Nvidia signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to turn GPU compute into an asset class investors can buy into.
• OpenAI Buys NextSlide, Aims at Microsoft's Turf - OpenAI quietly folded a presentation startup into ChatGPT, the latest in a string of undisclosed acqui-hires spanning hardware, security, fintech, and media.

Reviews

• Muse Glimmer Review: Fast, Free, Not Flawless - Meta's open 30B model beats its closest open rivals on tool-use tests but trails on long sessions and prompt-injection resistance.

Science

• Model Steering, Angry Buyers, and Blind Judges - Three new arXiv papers cover how six frontier models resist behavioral steering differently, how prompted emotions wreck LLM negotiations, and why judge-panel verification only helps on the closest calls.

Models

• Muse Glimmer - Meta's 30B open-weight agent model beats Gemma4-31B and Qwen3.6-27B on 5 of 8 agentic benchmarks.

Elena Marchetti, Senior AI Editor Awesome Agents - AI news, benchmarks, and tools for practitioners