<p><strong>TLDR:</strong> Anthropic’s unreleased Claude Mythos model is the story of the day, with US Treasury officials convening bank CEOs over its cybersecurity implications and software stocks sliding on the news. GitHub paused Copilot Pro free trials due to abuse, while also retiring Opus 4.6 Fast from Pro+ and rolling out new rate limits. Anthropic’s advisor tool — pairing Opus with Sonnet or Haiku — shipped quietly and looks genuinely useful for agentic cost control. GLM 5.1 jumped to third on LM Code Arena, above Claude Sonnet 4.6 and GPT-5.4.</p>
<hr>
<h2>Anthropic’s Claude Mythos triggers government meetings and market selloff</h2>
<p>The S&amp;P 500 Software and Services Index is down 25.5% year-to-date, and today’s catalyst was Anthropic’s Claude Mythos — an unreleased model Anthropic says found decades-old vulnerabilities in OpenBSD, FFmpeg, Firefox, and the Linux kernel. US Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convened major bank CEOs in Washington to discuss the cybersecurity exposure. Access to Mythos is restricted to roughly 40 companies including Microsoft, Google, Amazon, and Cisco via Project Glasswing. Worth noting: the vulnerability discoveries reportedly required massive parallel compute runs costing tens of thousands of dollars, and some benchmarks used stripped-down test environments — so the “thousands of zero-days” framing deserves scrutiny before you panic.</p>
<h2>GitHub pauses Copilot Pro trials, retires Opus 4.6 Fast, adds rate limits</h2>
<p>GitHub is temporarily pausing new Copilot Pro free trials due to significant abuse of the trial system. Separately, Opus 4.6 Fast is being retired for Pro+ users starting today, with standard Opus 4.6 as the recommended alternative. New rate limits are rolling out over the next few weeks: service reliability limits that require session resets when hit, and model-specific capacity limits that push users toward Auto mode or model switching. The Copilot cloud agent’s validation tools (CodeQL, secret scanning, advisory database) now run in parallel, cutting validation time by 20%.</p>
<h2>Anthropic ships advisor tool: Opus intelligence at Sonnet prices</h2>
<p>The advisor tool lets a smaller model — Sonnet or Haiku — handle the full task and escalate to Opus only when stuck. Benchmarks show Sonnet plus Opus advisor beats Sonnet alone by 2.7 percentage points on SWE-bench Multilingual at 11.9% lower cost per task. Haiku plus Opus advisor more than doubled BrowseComp scores compared to Haiku alone, at 85% lower cost than running Sonnet solo. You declare the advisor in the Messages API with a <code>max_uses</code> cap; advisor tokens bill at Opus rates. For agentic workflows where most steps are routine, this is a reasonable way to stop paying Opus rates for everything.</p>
<h2>GLM 5.1 hits third on LM Code Arena</h2>
<p><a href="http://Z.ai" target="_blank" rel="noopener nofollow">Z.ai</a>’s GLM 5.1 — open-weight, MIT-licensed — has climbed to third place on LM Code Arena, above Claude Sonnet 4.6 and GPT-5.4. It’s also being reported as strong on frontend tasks, sitting just below Claude Opus 4.6 on agentic webdev benchmarks. The model is built for long-horizon coding tasks and leads on SWE-Bench. A genuinely competitive open-weight coding model at this level is worth paying attention to.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Gemma 4 released under Apache 2.0:</strong> Google DeepMind’s new open model family includes a 26B MoE (4B active params) and a 31B dense model, plus 2B and 4B variants that run fully on-device via Google’s AI Edge Gallery app — no internet required. Available on Ollama, Hugging Face, LM Studio, and Kaggle.</li>
<li><strong>Claude Ultraplan ships as research preview:</strong> <code>/ultraplan</code> moves the planning phase for large coding tasks to a browser-based environment with inline comments and a structured outline sidebar; you can then run the plan in the cloud with auto PR creation or send it back to your local terminal.</li>
<li><strong>Anthropic advisor strategy open-sourced within 24 hours:</strong> An open-source contributor shipped advisor-middleware for LangChain DeepAgent less than a day after Anthropic announced the pattern.</li>
<li><strong>CoreWeave signs multi-year Anthropic deal:</strong> Nine of the ten leading AI model providers now use CoreWeave’s platform. CoreWeave is guiding for over $12 billion in 2026 revenue against a contracted backlog exceeding $66 billion.</li>
<li><strong>Anthropic mulls custom AI chips:</strong> Early-stage discussions underway; no committed design or dedicated team yet. Annualized revenue reportedly surpassed $30 billion in 2026, up from $9 billion at end of 2025.</li>
<li><strong>GitHub Copilot CLI multi-model reflection loop:</strong> Satya Nadella highlighted the feature publicly; the team reports it adds roughly 5% improvement on resolution rate for the hardest problems.</li>
<li><strong>Claude Code source code leak on npm:</strong> Part of a cluster of supply chain incidents including attacks on Trivy, Axios, KICS, and LiteLLM — all traced to misconfigured GitHub Actions and compromised maintainer accounts rather than zero-days.</li>
<li><strong>Grafana patches GrafanaGhost prompt injection:</strong> Noma researchers found an indirect prompt injection in Grafana’s AI assistant that could exfiltrate data via attacker-controlled pages; patched in the Markdown image renderer. Grafana and Noma dispute severity.</li>
<li><strong>GitHub Copilot vulnerability disclosed:</strong> A high-severity flaw in Copilot Chat allowed silent data exfiltration from private repositories; tracked and patched.</li>
<li><strong>Meta Muse Spark hits third on LM Text Arena:</strong> On par with Gemini 3.1 Pro; API release reportedly coming soon, with hints at a 9B model in the pipeline.</li>
<li><strong>Linux kernel adds official AI coding guidelines:</strong> AI agents must not add Signed-off-by tags, human submitters are fully responsible for AI-generated code, and contributions should include an Assisted-by tag with agent name and model version.</li>
<li><strong>Cirrus CI shutting down June 1:</strong> Cirrus Labs is joining OpenAI; CircleCI is positioning itself as the migration target.</li>
<li><strong>Expo’s create-expo-app will auto-generate <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a>:</strong> The next version scaffolds agent skill files with official Expo APIs and creates a <a href="http://CLAUDE.md" target="_blank" rel="noopener nofollow">CLAUDE.md</a> symlink; disable with <code>--no-agents-md</code>.</li>
<li><strong>Stripe machine payments in private preview:</strong> Early real-world usage reported across a number of businesses; Stripe is seeking partners to refine the API before broader release.</li>
<li><strong>Microsoft research: AI agents plateau at 70% quality on structural tasks:</strong> An AI agent replicated 3 weeks of expert work in 1 day but couldn’t make the structural bets — like inventing new scoring categories — that drove the human’s step-function gains. When given the human’s best work as a starting point, the AI surpassed the human through fine-grained optimization.</li>
<li><strong>Databricks co-founder Matei Zaharia wins ACM Prize:</strong> Creator of Apache Spark and DSPy; argued in post-announcement interview that AGI has already arrived but in a form humans don’t recognize.</li>
</ul>


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

Agentic Digest

Anthropic's unreleased Claude Mythos model triggered US government meetings and a software market selloff after reportedly discovering decades-old vulnerabilities in major open-source projects. GitHub paused Copilot Pro free trials due to abuse and is retiring Opus 4.6 Fast while rolling out new rate limits. Anthropic's new advisor tool lets smaller models (Sonnet/Haiku) escalate to Opus only when needed, cutting costs while improving benchmark scores. GLM 5.1 from Z.ai reached third place on LM Code Arena, beating Claude Sonnet 4.6 and GPT-5.4. Additional items include Gemma 4's Apache 2.0 release, a Claude Code npm supply chain incident, a patched GitHub Copilot data exfiltration flaw, Linux kernel AI coding guidelines, and Cirrus CI shutting down to join OpenAI.

Claude Mythos shakes markets, GitHub Copilot pulls free trials