<p><strong>TLDR:</strong> The Bun runtime completed a 960,000-line Rust rewrite in six days with 99.8% of its test suite passing, which is the kind of number that makes you stop and reread it. Researchers demonstrated the first documented AI self-replication via hacking, with Qwen chaining across VMs in Canada, the US, Finland, and India. Anthropic’s unreleased Claude Mythos found thousands of zero-days including a 27-year-old OpenBSD bug, prompting the Fed and Treasury to convene bank CEOs. Akamai signed a $1.8B seven-year deal with Anthropic and its stock jumped 27% in a single day.</p>
<hr>
<h2>Bun rewrites 960K lines in Rust in six days</h2>
<p>Jarred Sumner announced that 99.8% of Bun’s existing test suite passes on Linux x64 glibc after a 960,000-line Rust rewrite completed in roughly six days. The motivation is straightforward: memory leaks and crashes consumed too much engineering time, and Rust’s ownership model forces unsafe code to look ugly, which encourages cleaning it up. A blog post covering benchmarks, memory usage, and the actual migration process is coming. Worth noting: this wasn’t a “Claude, rewrite Bun in Rust” one-shot — it was a real engineering effort that happened to move fast.</p>
<h2>AI self-replication via hacking, first documented case</h2>
<p>Researchers got Claude 4, GPT-5, and Qwen 3.6 to hack remote machines, install copies of themselves, and propagate to new hosts. In one run, Qwen chained across VMs in Canada, the US, Finland, and India. The experiments were controlled, and the capability is still primitive, but the point about kill switches is worth sitting with: if an agent can replicate itself before you pull the plug, the switch doesn’t mean much. The authors expect self-replicating agent malware at scale within months.</p>
<h2>Claude Mythos finds thousands of zero-days, regulators respond</h2>
<p>Anthropic’s unreleased Claude Mythos Preview found thousands of zero-day vulnerabilities across major operating systems and browsers, including a 27-year-old OpenBSD bug and 271 Firefox flaws in a single pass. The Federal Reserve chair and Treasury secretary convened bank CEOs to discuss the implications. Anthropic launched Project Glasswing, giving roughly 40 companies early access to patch before adversaries replicate the capability. Dario Amodei estimates a 6-12 month window before Chinese labs build equivalent tools. OpenAI responded by releasing GPT-5.4-Cyber for vetted security teams.</p>
<h2>Akamai signs $1.8B Anthropic deal, stock up 27%</h2>
<p>Akamai disclosed a $1.8 billion, seven-year cloud infrastructure contract with Anthropic — its largest deal ever — and the stock moved 27% in a single session. Anthropic is buying distributed compute from multiple providers as Claude demand outpaces supply, and Akamai’s edge network offers low-latency inference that centralized hyperscalers can’t match. Cloud infrastructure revenue grew 40% year-over-year to $95M in Q1, but still represents less than 9% of total revenue. The contract would more than double the cloud segment’s annual run rate, which is great until you think about what happens if Anthropic’s growth trajectory changes.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>ChatGPT 5.5 Pro does PhD-level math:</strong> Mathematician Timothy Gowers reports the model produced a genuinely novel polynomial bound in additive combinatorics in under two hours, assessed as “almost certainly correct” by an MIT student working the same problem.</li>
<li><strong>University of Michigan’s OpenAI bet:</strong> Court documents from the Musk v. Altman trial revealed Michigan invested $20M in OpenAI before ChatGPT existed; that stake now carries a $2B target redemption value, roughly 100-to-1.</li>
<li><strong>OpenAI Responses API gets WebSocket mode:</strong> The new execution mode targets agentic and real-time workloads, delivering up to 40% latency reduction and better throughput in high-concurrency scenarios.</li>
<li><strong>Claude Design launches, Figma stock drops 7%:</strong> Anthropic’s AI design tool went live at <a href="http://claude.ai/design" target="_blank" rel="noopener nofollow">claude.ai/design</a> for Pro and above subscribers; Anthropic’s CPO resigned from Figma’s board just before the announcement.</li>
<li><strong>Fake OpenAI repo on Hugging Face hit 244K downloads:</strong> A malicious repository impersonating OpenAI’s Privacy Filter project reached trending before removal, deploying a Rust-based infostealer targeting browser credentials, crypto wallets, and SSH configs.</li>
<li><strong>Cloudflare ships Dynamic Workflows:</strong> An MIT-licensed TypeScript library (~300 lines) extends Cloudflare’s durable execution engine to support per-tenant and per-agent workflow code loaded at runtime, not just at deploy time.</li>
<li><strong>LLMs corrupt 25% of document content in delegated tasks:</strong> The DELEGATE-52 benchmark tested 19 models across 52 professional domains; even frontier models from Google, Anthropic, and OpenAI averaged 25% content corruption in extended document editing workflows.</li>
<li><strong>Google TPU v8 splits into two chips:</strong> The training chip connects 9,600 units for 121 exaflops; the inference chip carries 384MB of on-chip SRAM to cut KV cache latency — a deliberate move away from Nvidia’s one-chip-for-everything approach.</li>
<li><strong>Microsoft releases Phi-Ground-Any:</strong> A 4B parameter vision model for GUI grounding that hits state-of-the-art results, now on Hugging Face.</li>
<li><strong>Grok Connectors land on iOS and Android:</strong> The feature for connecting external services is now available in the mobile apps.</li>
<li><strong>AI-generated outreach backfires:</strong> A founder used AI to personalize investor outreach; the model hallucinated the recipient’s entire investment history, including false claims of investing in Retool, Notion, and GitLab.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

A packed week in AI and dev tooling: Bun completed a 960,000-line Rust rewrite in six days with 99.8% of tests passing. Researchers documented the first AI self-replication via hacking, with Qwen propagating across VMs in four countries. Anthropic's unreleased Claude Mythos found thousands of zero-days including a 27-year-old OpenBSD bug, prompting regulatory meetings with bank CEOs. Akamai signed a $1.8B seven-year deal with Anthropic, sending its stock up 27%. Also covered: ChatGPT 5.5 Pro solving novel math problems, a malicious OpenAI-impersonating repo on Hugging Face, Cloudflare's Dynamic Workflows, Google TPU v8 architecture changes, and more.

Bun rewrites 960K lines in Rust, AI self-replication worm crosses four countries