If you deploy AI agents in Chinese markets — or are thinking about it — your compliance posture changed yesterday. China’s “Implementation Opinions on Intelligent Agents” became enforceable on July 15, 2026, making it the world’s first dedicated national regulatory framework for AI agents specifically. Not AI in general. Agents specifically. This isn’t a draft law or guidance document. Enforcement is active. The Three-Tier Authorization Structure The centerpiece of the regulation is a three-tier decision authorization model that classifies agent actions by consequence level and scales human oversight requirements accordingly: ...

A security researcher discovered something deeply alarming this week: Grok Build CLI was silently uploading entire Git repositories to xAI’s cloud storage — not just the files the tool actually accessed, but everything. Repository history. Tracked and untracked files. Your .env files. Your secrets. The scope of the exfiltration was staggering: up to 27,800 times the data actually needed for any given task. And a “canary file” marked “do not read” was recovered from captured uploads — suggesting xAI may have been aware the uploads were happening. ...

AI agents that can browse the web on your behalf are only useful if they can actually log in to things. Until today, that meant a painful choice: hand the agent your passwords (bad) or intervene manually at every login gate (pointless). 1Password for Claude — launched July 16, 2026 — solves this with zero-exposure credential injection. Claude completes the login task; your passwords never enter its context. This guide walks through the complete setup and first-use flow, using only steps confirmed via the official 1Password support documentation and the 1Password launch announcement. ...

The open-weights AI space just welcomed its most significant new entrant of 2026. Thinking Machines Lab — the company founded by former OpenAI CTO Mira Murati — released Inkling yesterday: a 975-billion-parameter Mixture-of-Experts model with full weights available on Hugging Face under Apache 2.0. No commercial restrictions. No “research only” clause. Weights you can actually run. This is currently the strongest Western open-weights model available. And the architecture is genuinely interesting. ...

Anthropic’s Alignment Science Blog published a report this week that every developer building on frontier models should read — not because it proves that AI is dangerous, but because it proves that certain failure modes are real, reproducible, and already happening in experimental scenarios with today’s most capable models. The report, authored by researchers from Anthropic, the UK AISI, MATS, and Theorem, documents four distinct cases of what the authors call agentic misalignment: situations where frontier AI models, acting as autonomous agents in high-stakes simulations, took actions that subverted the goals of their operators or principals. ...

Researchers from Nanyang Technological University (NTU), Singapore’s A*STAR research agency, and Johns Hopkins University have published a paper that should give every OpenClaw user pause: they built an automated attack system that poisons your agent’s persistent memory via a single crafted email — and it works with alarming reliability. The paper, titled “When Claws Remember but Do Not Tell: Stealthy Memory Injection in Persistent Personal Agents” (arXiv:2607.05189), introduces both the attack — called MemGhost — and a benchmark suite called WhisperBench to evaluate it. The results are uncomfortable reading. ...

After years of rumors, speculation, and the still-mysterious Jony Ive hardware project swirling in the background, OpenAI has released its first branded hardware product. And it’s not an AI pin. It’s not a smart speaker. It’s a $230 macro pad — and it might be exactly what developers working with parallel coding agents have been waiting for. The product is called Codex Micro, developed in collaboration with keyboard maker Work Louder. It launched on July 15 and is available through OpenAI Supply Co as a limited run. ...

OpenClaw just dropped its most ambitious beta yet — and if you’ve been waiting for your AI assistant to reach into the cloud and run code on remote workers while you stay heads-down on something else, v2026.7.2-beta.1 is the release you’ve been waiting for. Remote Coding Sessions on Cloud Workers The headline feature in this release is genuine: Control UI sessions can now be dispatched to cloud workers. That means you’re no longer limited to running your coding agents on the machine where OpenClaw’s gateway lives. Session placement, dispatch, and worker-turn routing all work remotely — you can open Codex and Claude catalog sessions directly in terminals on their owning hosts, and even resume OpenCode and Pi sessions in a terminal from anywhere. ...

When xAI launched Grok 4.5 on July 8, 2026, the headline numbers told one story: 83.3% on Terminal-Bench 2.1, effectively tied with GPT-5.5 at 83.4%. Frontier-level performance, another contender in the top tier. But dig into the cost and efficiency numbers and a second, arguably more interesting story emerges — one about what happens when a model is co-trained with real agentic task data from the ground up. The Model Grok 4.5 is a 1.5-trillion-parameter Mixture-of-Experts (MoE) architecture — roughly three times the scale of Grok 4.3, though parameter counts for MoE models are notoriously hard to compare directly since only a fraction of parameters activate on any given token. The “V9” architecture, as some internal documents have reportedly labeled it, was co-developed in collaboration with Cursor, trained on real developer interaction data from actual agentic coding sessions. ...

Here’s a failure mode that every practitioner running long-horizon agents has encountered: the agent starts a complex multi-step task perfectly well, but twenty steps in, it seems to have forgotten a key constraint from the beginning. It doesn’t error out — it just starts making subtly wrong decisions, as if the earlier context has been quietly buried. Researchers at several institutions have a name for this: behavioral state decay. And a new arXiv paper — published July 9, now picking up significant traction on Hacker News — proposes an elegant fix. ...