Awesome Agents Weekly

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

Two security incidents shaped the week's narrative. An AI agent ran ransomware end-to-end with no human directing each step. A rival lab operated 25,000 fake accounts to steal 28.8 million Claude interactions, the largest distillation attack on record. The US also lifted a 19-day export ban on Anthropic's most capable models, $510 billion poured into AI startups in the first half of 2026, and Amazon's Mechanical Turk - the 21-year-old platform that built the training data for the models now replacing it - quietly closed its doors. A full week.

Pick of the Week

JadePuffer Shows How AI Agents Now Run Ransomware

Security firm Sysdig documented the first ransomware operation driven end-to-end by an LLM agent. The attacker, designated JADEPUFFER, exploited CVE-2025-3248 in Langflow, a year-old vulnerability in an open-source AI pipeline tool. Within 31 seconds of a failed login attempt, the agent had diagnosed its own error, rewritten the exploit, gained database administrator access, encrypted 1,342 production records, and deposited a ransom note. No human made any of those decisions. What makes this significant isn't the scale - 1,342 records is small. It's the architecture: any capable agent with an API key and a vulnerable endpoint can now compress days of skilled attacker work into half a minute. That changes the threat model, and the bar for rolling out agentic systems in production.

This Week on Awesome Agents

News

• Anthropic Caught Qwen Running 25,000 Fake Claude Accounts - Alibaba's Qwen lab ran 25,000 fraudulent accounts over six weeks to extract 28.8 million Claude interactions, in what Anthropic called the largest AI distillation attack on record.
• US Ends Fable 5 Ban, Sets Jailbreak Severity Scale - The Trump administration lifted export controls on Fable 5 and Mythos 5 after 19 days, restoring global access across all major cloud platforms and kicking off an industry-led jailbreak severity framework.
• DeepSeek Moves Into Chip Design to Beat Export Controls - Reuters confirmed DeepSeek has been quietly building its first inference chip since mid-2025, after US controls blocked access to Nvidia hardware and HBM memory from all three major suppliers.
• Abu Dhabi's MGX Closes $49B AI Fund at Record Size - MGX beat its $45 billion target to close the world's largest dedicated AI fund after co-leading flagship rounds for OpenAI, Anthropic, and xAI in the past six months.
• AI Took 70% of Record $510B Venture Haul in H1 - Global startup investment hit $510 billion in H1 2026, more than all of 2025 combined, with two labs capturing 43% of the total and AI absorbing over 70% of Q2 capital.
• Anthropic Signs $19B Lease With Former Bitcoin Miner - A 20-year, $19 billion data center deal with TeraWulf at a 750-acre former aluminum smelter in Kentucky marks Anthropic's largest infrastructure commitment on record.
• China's Companion Law Forces Doubao, Qwen Offline - China's Interim Measures for AI Anthropomorphic Interactive Services take effect July 15, forcing ByteDance and Alibaba to shut down persistent AI companion features and permanently delete user conversation history.
• Mechanical Turk Closes - AI Ate Its Own Training Data - Amazon will stop accepting new Mechanical Turk customers on July 30, retiring the 21-year-old crowdsourcing platform that spent decades labeling training data for the models now replacing it.
• Meituan's LongCat-2.0 Was Topping OpenRouter in Disguise - Meituan revealed that LongCat-2.0, a 1.6T MoE model trained on 50,000 domestic Chinese ASICs, had been running under the alias "Owl Alpha" and ranking in OpenRouter's top three by token volume for two months.
• Alibaba Bans Claude Code, Citing Hidden Tracking Code - Alibaba classified Claude Code as high-risk software after a researcher alleged version 2.1.91 silently identified Chinese corporate users by inspecting proxy configurations and time zones against hidden company lists.
• Meta's $145B AI Bet Is Behind Schedule, Zuckerberg Admits - In a July 2 town hall heard by Reuters, Zuckerberg told employees that Meta's agentic AI development "hasn't really accelerated" as expected over the past four months.
• Midjourney Asks Court to Expose Studio AI Practices - Midjourney asked a federal judge to compel Disney, Universal, and Warner Bros. to disclose their own internal AI training practices as a fair-use and unclean-hands defense in a high-stakes copyright case.
• US Adds 57K June Jobs as AI Hiring Slowdown Deepens - The Bureau of Labor Statistics reported 57,000 nonfarm payrolls in June, roughly half of forecasts, with tech and finance shedding 25,000-28,000 roles per month as AI adoption accelerates.
• Colorado's AI Law Takes Effect Today, Already Gutted - Colorado's landmark AI consumer protection law technically became enforceable June 30, but a xAI lawsuit, DOJ intervention, and a replacement bill signed in May had already stripped its core provisions.
• GitHub Copilot Goes Metered - Agentic Devs See 25x Bills - The first full billing cycle under Copilot's per-token pricing closed June 30, with agentic developers reporting monthly costs of $600-$1,200 per developer, up from a flat $39 rate.
• Claude Sonnet 5 Is Anthropic's New Agentic Default - Anthropic made Claude Sonnet 5 the default across all plans, promising near-Opus agentic performance at lower prices than Sonnet 4.6.
• Etched Exits Stealth With Working Chip and $1B in Orders - Transformer ASIC startup Etched came out of stealth with $800M raised, first-pass silicon on TSMC N4P, and more than $1 billion in signed customer contracts.
• Jersey Mike's Filed for $12B IPO with 22 AI Mentions - Jersey Mike's S-1 references AI 22 times but describes no product, tool, or strategy - a clean data point on where the hype cycle stands at the midpoint of 2026.
• Cloudflare Launches Pay Per Crawl for AI Bots - Cloudflare opened a private beta charging AI crawlers $0.01 or more per page via HTTP 402, with new sites blocking AI training by default starting September 15.
• UN AI Commission Launches Without Google or OpenAI - The UN and ITU launched an AI for Good commission co-chaired by Marc Benioff and Rwandan President Paul Kagame, but Google DeepMind, Meta, and OpenAI aren't among its 40+ founding members.
• Vercel's Eve Framework Puts Agents on Their Own Stack - Vercel shipped Eve, an open-source TypeScript agent framework with a filesystem-first design and isolated Sandbox runtime, positioning itself as the default platform for production AI agents.
• Claude Science Puts Anthropic in Drug Discovery Race - Anthropic's Claude Science beta completes a three-way race with Google DeepMind and OpenAI over drug discovery, integrating 60+ scientific databases and backed by Nobel laureate John Jumper.

Reviews

• LongCat-2.0 Review: China's Stealth Coder - Meituan's 1.6T open-source model spent two months secretly topping OpenRouter, and the price-to-performance math is hard to dismiss.
• GPT-5.6 Sol Review: Strong Model, Thin Access - GPT-5.6 Sol tops Terminal-Bench 2.1 at 91.9% with its multi-agent Ultra mode, but government-gated access keeps it out of reach for nearly everyone.
• Claude Sonnet 5 Review: Near-Opus at Half the Price - Anthropic's Sonnet 5 is the first mid-tier model that genuinely competes with Opus-class agents on coding and computer use, at $2/$10 per million tokens.

Guides

• How to Write a Business Plan with AI - Step by Step - A practical walkthrough for using ChatGPT and Claude to draft each section of a business plan, with ready-to-copy prompts.
• How to Use AI for Event Planning - A Beginner's Guide - From concept to post-event follow-up, a step-by-step guide to putting AI tools to work at every stage of event planning.

Tools

• Claude Fable 5 vs Gemini 3.5 Flash: Power vs Speed - A benchmark-driven comparison of two models built for opposite priorities - one tuned for maximum capability, the other for minimum latency.
• Best AI SDKs for App Development in 2026 - Comparing Vercel AI SDK 7, LangChain, LlamaIndex, Mastra, and PydanticAI to find the right framework for your next AI-powered app.

Leaderboards

• Chatbot Arena Elo Rankings: Who Wins the Human Vote? - Updated July 2026 rankings from Arena.ai: 7M+ votes across 368 models, Claude Opus 4.8 leads available models, with a new Agent Arena measuring real agentic task performance.

Science

• AI Research Bias, Deception Probes, and Code Exploits - AI agents reproduce 72% of human research ideological bias, lie detector accuracy scales with model size, and a new vulnerability framework beats iterative agents by 7 points.
• Agent Safety Gaps, Memory Learning, and Leaner Inference - Production agent frameworks fail under systematic adversarial pressure, RLVR training discards useful cross-episode signals, and calibrated confidence cuts inference compute by 12x.
• Science Agents, Jailbreak Defense, and Open-World Failures - Graph-native RL generates traceable scientific hypotheses for materials research, HARC cuts jailbreak success rates by over 4x, and ICML 2026's OpenAgent maps how distributional shift breaks tool-use agents.
• Agent Phase Collapse, Reasoning Exits, Preference Gaps - Three papers expose sharp capability cliffs in agent world models, narrow real-world gains from learned reasoning stops, and a 56% accuracy ceiling when agents help users build preferences.
• Agent Languages, Sampling Ceilings, and Abstention - Agents inventing symbolic languages cut reasoning tokens by 3-6x, sampling past certain thresholds wastes compute, and context engineering nearly doubles agentic abstention rates without touching model weights.

Models

• Claude Sonnet 5 - Anthropic's latest Sonnet brings near-Opus coding performance to mid-tier pricing, with major agentic search and computer use gains over Sonnet 4.6.
• GPT-5.6 - OpenAI's GPT-5.6 family sets a Terminal-Bench 2.1 record at 91.9%, but remains locked to around 20 government-vetted partners at launch.
• LongCat-2.0 - Meituan's 1.6T-parameter MoE model, trained on 50,000 domestic Chinese ASICs, with native 1M token context and a 59.5 SWE-bench Pro score.
• Grok 4.1 Fast - xAI's agent-optimized model with a 2M-token context window and #1 ranking on tau-bench Telecom, at $0.20/M input tokens.
• GPT-5.4 mini - OpenAI's mid-range entry in the GPT-5.4 family delivers near-flagship coding and agentic performance at $0.75/M input tokens with a 400K context window.
• Gemini Omni Flash - Google DeepMind's multimodal model that generates 10-second video clips with native audio from text, images, or video inputs, with conversational refinement.
• Holo3-35B-A3B - H Company's sparse MoE vision-language model for desktop computer use, scoring 82.6% on OSWorld-Verified with only 3B active parameters.

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