6 min read · 14 sources

Researchers chained two critical vulnerabilities to compromise multiple OpenAI employees’ ChatGPT and Codex accounts in July, and they were inside the company’s internal repositories in under 72 hours. The attack went through OpenAI’s Discourse forum at community.openai.com, not through ChatGPT itself. They proved access by opening a harmless PR in OpenAI’s internal monorepo and got paid a $6,500 bounty for it. The full writeup is a masterclass in how third-party infrastructure becomes the soft underbelly of a company’s identity and code.

That’s the biggest story of the day, but it’s far from the only one. Anthropic says Claude now drives 26% of its own R&D. Huawei is shipping AI systems to hundreds of Chinese customers while Nvidia still leads on process tech. And Safari quietly shipped an MCP server that turns your coding agent into a browser automation tool. Here’s the rest.

A 13-minute full test suite now blocks merges, requiring a merge queue even for a two-person project.

OpenAI Got Hacked Through Its Own Forum

Source: hacktron.ai ↗

The HacktronAI writeup walks through a two-vulnerability chain that started on OpenAI’s Discourse instance. The researchers achieved remote code execution and administrative access on the forum, then pivoted to employee accounts for ChatGPT and Codex. From there they reached internal OpenAI repositories and potentially connected services like GitHub and Slack.

The full timeline from discovery to repo access was under 72 hours. The team proved the access by opening a harmless PR in OpenAI’s internal monorepo and collected a $6,500 bounty.

The lesson for everyone running a SaaS company: your community forum is an attack surface with the same trust level as your production identity infrastructure. Account takeovers via third-party infrastructure are fast, cross-product, and don’t require a single exploit against your core product.

Anthropic Says Claude Does 26% of Its R&D

Source: bloomberg.com ↗

Anthropic is claiming that Claude drives more than a quarter of its research and development work. There are over 30,000 agents doing research and engineering work at Anthropic at any one time, and staffers collaborate with the chatbot for about 90% of their work.

The company is building a framework to track agents and monitor how much work they do. That’s the part to watch: nobody has a good accounting system for agent-contributed labor yet, and Anthropic is trying to build one for itself while selling the same tools to everyone else. Whether the 26% number is measurement or marketing, the direction is clear.

Huawei's Play to Be China's Nvidia

Source: wsj.com ↗

Huawei plans to release two new AI chips next year, and it has already shipped more than 1,000 AI-computing systems to over 370 customers. China’s AI chip market is projected to grow to $67 billion by 2030.

Huawei’s chip-making technology still trails Nvidia’s. The company is compensating with workarounds and new designs that squeeze in more computing power using less advanced equipment. That’s a systems-level engineering problem, not just a fab problem: if you can’t shrink the transistor, you change the architecture. Expect the two new chips to prioritize memory bandwidth and interconnect over raw process-node bragging rights.

Helix 2.5: Robots That Generalize to Homes They've Never Seen

Source: figure.ai ↗

Figure’s Helix 2.5 is a neural network for humanoid robots that achieves zero-shot whole-body autonomy across 30 unseen Bay Area homes for three long-horizon behaviors: tidying, towel folding, and bed making.

The numbers matter. Pretrained on Figure’s Index dataset, the model improved zero-shot success from 9% to 56% compared to a baseline without Index pretraining, while using half the task-specific data of Helix 02. Figure also reports a scaling law where doubling Index pretraining data improves downstream robot-action prediction, allowing loss forecasting to four decimal places.

For engineers, this is the first credible evidence that generalizable robot manipulation doesn’t require environment-specific fine-tuning. Index is now generating roughly 35 minutes of new human experience every second, which is the data engine behind that scaling law.

Raptor 3: How SpaceX Cut the Pipes

Source: construction-physics.com ↗

The Raptor engine’s evolution from Raptor 1 to the Raptor 3 that first flew in May is stark. The redesign consolidated pipes and wires to the point where many people initially didn’t believe SpaceX’s photo of the three generations. The streamlining also brought about 35% more thrust than the original.

SpaceX publishes no official schematics, so the details are inferred from Musk’s comments and fan speculation. The engine uses full-flow staged combustion, and the article explains how the operating principles and performance gains fit together. For anyone designing complex systems, the lesson is that iterative simplification is a performance feature, not just a maintenance convenience.

Safari 27.0 Ships an MCP Server for Browsers

Source: webkit.org ↗

Safari 27.0 introduces Safari MCP, a Model Context Protocol server that lets coding agents like Claude Code or Codex control the browser to inspect DOM, network requests, screenshots, and console output. It runs locally with no network calls and no access to personal Safari data, enabled via Safari > Settings > Developer > “Allow remote automation and external agents.”

That turns your browser into a test harness for coding agents. The release also includes customizable <select> elements, scroll anchoring, the <model> element for 3D on iOS/iPadOS/macOS, and visionOS immersive environments. The release notes are the longest yet at 83 features.

Crusoe Swaps Megascale for Flatbed Trucks

Source: wsj.com ↗

Crusoe built one of the world’s biggest AI supercomputers in Abilene, Texas. Now it’s investing heavily in smaller data centers, manufacturing them in factories, loading them onto flatbed trucks, and deploying them wherever power is available.

The logic is sound: giant clusters with hundreds of thousands of chips are useful for training, but inference can run on considerably smaller amounts of chips. Factory-built, truck-delivered data centers are faster and cheaper than large construction projects. That’s the same playbook as modular nuclear and microgrids: go where the power is, don’t wait for the grid to come to you.

The Test Suite Is the New Code Review

Source: allen.hutchison.org ↗

With AI agents opening and reviewing pull requests, the test suite has become the bottleneck in the software pipeline. The author describes how a 13-minute full test suite now blocks merges, requiring a merge queue even for a two-person project, and that tiered testing can’t catch bugs before production picks up changes on a 15-minute cycle.

Code review used to be slow, but agentic review took that bottleneck away almost entirely. Now the tests are the critical gate, which means CI/CD design has to change: tests must be fast and comprehensive enough to serve as the primary safety net, because a human reviewer isn’t going to catch what the agents miss.

Quick Hits

Source: techcrunch.com ↗

• Automattic’s $8.15M ouster bill: During Matt Mullenweg’s 33-hour paid leave, interim CEO Mark Davies and CLO Andy Missan signed reciprocal severance deals - 12 months of base salary, accelerated equity vesting, and a year of health coverage each. Mullenweg fired them on return; the legal team now decides whether to pay or fight the “cause” definitions.
• Bitcoin’s quantum migration debate: Legacy ECDSA and Schnorr signatures using secp256k1 could become vulnerable to quantum attacks. Every mitigation - doing nothing, freezing vulnerable coins, or forcing migration - violates some Bitcoin principle like property rights or user sovereignty.
• Two techniques for System One models: Tiered goals and tournament choice sampling let you build fast, decision-only classifiers from existing LLMs by batching single-token structured-output prompts, no architectural changes needed.
• Bend: A new language for the post-AGI economy that targets massive parallelism without threads or locks, with compiler-enforced rules that block AI-generated code violating declared invariants.
• How to write with an LLM: Two rules from sockpuppet.org: never use a single word an LLM suggests, and avoid encouragement from the model. Write it yourself, then use the model as a copyeditor.