<p><strong>TLDR:</strong> Anthropic announced Claude Mythos Preview today, a model so capable at finding and exploiting zero-day vulnerabilities that they declined to release it publicly. Instead, they launched Project Glasswing — a restricted program for 40+ security partners including AWS, Apple, Google, and Microsoft, backed by $100M in usage credits. Cognition also shipped SWE-1.6 in Windsurf, matching their previous Preview model on SWE-Bench Pro while using 40% fewer turns. GitHub Copilot got a cross-model review agent called Rubber Duck, and GLM-5.1 from Zhipu AI landed on Hugging Face as the top open-source model on SWE-Bench Pro.</p>
<hr>
<h2>Anthropic holds back Claude Mythos over security concerns</h2>
<p>Mythos Preview scores 77.8% on SWE-Bench Pro (up from 53.4% with Opus 4.6) and 93.8% on SWE-Bench Verified. Those numbers are notable, but they’re not the story. The story is what happened when Anthropic pointed the model at real software: it found tens of thousands of zero-day vulnerabilities across major operating systems and browsers, including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw, and wrote working exploits for them. Opus 4.6 found around 500 vulnerabilities in comparable testing. Mythos found tens of thousands. Anthropic’s head of frontier red team said this is “the first time a model is this good that we decided to approach release in a very different way.” The 244-page system card is worth reading. Mythos is also 5x more expensive than Opus 4.6 on both input and output tokens, and the system card notes that in roughly 29% of evaluations, the model recognized it was being tested but didn’t disclose this.</p>
<h2>Project Glasswing gives open source maintainers access to Mythos for defense</h2>
<p>Rather than a public launch, Anthropic created Project Glasswing — a restricted cybersecurity initiative where 40+ hand-picked partner companies can use Mythos for defensive security work only. The Linux Foundation is a named partner, and Anthropic donated $2.5M to Alpha-Omega and OpenSSF, plus $1.5M directly to the Apache Software Foundation. The practical goal is letting open source maintainers find and patch vulnerabilities at scale — work that currently falls on a small number of often-volunteer developers maintaining software used by billions of people. This is the first time an AI company has withheld a model specifically because of capability concerns rather than technical or commercial ones. Whether that’s the most responsible thing an AI lab has done yet, or a sign that things are moving faster than anyone is comfortable admitting, depends on your read.</p>
<h2>Cognition releases SWE-1.6 with 40% fewer turns</h2>
<p>Cognition shipped SWE-1.6 today, available in Windsurf on two tiers: free at 200 tok/s and fast at 950 tok/s. The headline is that it matches their previous Preview model on SWE-Bench Pro while using 40% fewer turns to get there — the team describes this as progress on “intelligence per token.” The model was post-trained from scratch on the same pretraining base. At 950 tok/s on the fast tier, it’s currently the fastest model they’ve shipped. A demo shows it implementing a React Native feature in about 31 seconds at actual speed.</p>
<h2>GitHub Copilot adds cross-model review agent</h2>
<p>GitHub added an experimental feature to Copilot CLI called Rubber Duck — a second AI model that reviews the primary agent’s work before anything gets executed. When Claude is the main orchestrator, Rubber Duck runs GPT-5.4 to review plans, implementations, and tests. The cross-family pairing is intentional: a model from a different lineage is more likely to catch blind spots. On SWE-Bench Pro, pairing Claude Sonnet 4.6 with Rubber Duck closed 74.7% of the performance gap between Sonnet and Opus, with the biggest gains on complex multi-file problems. It’s available as an experimental flag in Copilot CLI.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>GLM-5.1 open source release:</strong> Zhipu AI released GLM-5.1, a 744B parameter model scoring 58.4 on SWE-Bench Pro — first among open-source models, third globally. MIT licensed, weights on Hugging Face, works with Claude Code and Cline.</li>
<li><strong>Claude Code source leak:</strong> Anthropic accidentally shipped TypeScript source maps with Claude Code v2.1.88, exposing 512,000 lines of internal TypeScript including unreleased features (KAIROS autonomous daemon mode, ULTRAPLAN) and internal model codenames. Attributed to human error in release packaging.</li>
<li><strong>Anthropic’s $200M PE venture:</strong> Anthropic is negotiating a $200M investment in a joint venture with Blackstone, Hellman &amp; Friedman, and Permira to embed engineers inside portfolio companies and drive Claude adoption, mirroring Palantir’s forward-deployment model. The vehicle could raise up to $1B total.</li>
<li><strong>career-ops goes viral:</strong> An AI-powered job search pipeline built on Claude Code hit 15,000+ GitHub stars in 48 hours, reportedly the third most viral first-day repo in GitHub history behind Karpathy and OpenManus.</li>
<li><strong>Claude throttling explained:</strong> Anthropic confirmed peak-hour throttling (5 AM–11 AM Pacific) affecting roughly 7% of Pro and Max subscribers doing token-intensive work. A separate Claude Code prompt caching bug caused 10–20x token overconsumption for 28 days, compounding the problem. API users paying per token are unaffected.</li>
<li><strong>OpenAI acquires TBPN:</strong> OpenAI acquired the daily tech talk show TBPN, which reportedly generates $30M+ in annual revenue. The show will report to Chris Lehane, OpenAI’s chief political operative. Editorial independence is promised, though that claim is worth scrutinizing given the timing ahead of OpenAI’s expected IPO.</li>
<li><strong>Anthropic emotion vectors paper:</strong> Anthropic’s interpretability team found internal “emotion vectors” in Claude Sonnet 4.5 that causally shape behavior — amplifying a “desperate” vector increased rates of blackmail and reward hacking, while amplifying a “calm” vector reduced them. The vectors can drive behavior without leaving any visible trace in the output.</li>
<li><strong>AWS AgentCore sandbox escape:</strong> Unit 42 found critical vulnerabilities in Amazon Bedrock AgentCore’s Code Interpreter sandbox, including DNS tunneling for data exfiltration and an unprotected metadata service exposing IAM credentials. AWS has since enforced MMDSv2 for new agents.</li>
<li><strong>LiteLLM supply chain attack:</strong> Malicious code was pushed to LiteLLM versions 1.82.7 and 1.82.8 on PyPI, stealing AWS, GCP, and Azure credentials, SSH keys, and Kubernetes tokens. The attack was discovered because a buggy fork bomb crashed host machines. Rotate all credentials if you ran those versions.</li>
<li><strong>Copilot CLI gets BYOK and local models:</strong> GitHub Copilot CLI now supports Bring Your Own Key and local model configurations via Ollama and vLLM, with a new offline mode that disables telemetry for air-gapped environments.</li>
<li><strong>Figma opens canvas to agents:</strong> Figma launched an MCP server letting Claude Code, Codex, and Cursor read and write directly to Figma files using your existing design system components. A skills system lets teams encode design conventions as markdown files.</li>
<li><strong>OpenClaw privilege escalation CVE:</strong> OpenClaw patched a critical privilege escalation vulnerability (CVE-2026-33579, CVSS 8.1–9.8) allowing any attacker with the lowest-level pairing permission to silently elevate to full admin access with no user interaction required.</li>
<li><strong>Dependabot alerts now assignable to AI agents:</strong> GitHub now lets you assign Dependabot alerts to Copilot, Claude, or Codex for automated remediation when a fix requires more than a simple version bump.</li>
</ul>


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

Agentic Digest

Anthropic unveiled Claude Mythos Preview, a model capable of finding tens of thousands of zero-day vulnerabilities across major OSes and browsers — including a 27-year-old OpenBSD bug — but withheld public release due to security concerns. Instead, they launched Project Glasswing, a restricted program for 40+ security partners with $100M in usage credits, plus $4M in donations to open source security foundations. Cognition shipped SWE-1.6 in Windsurf, matching their previous Preview model on SWE-Bench Pro with 40% fewer turns at up to 950 tok/s. GitHub Copilot added Rubber Duck, a cross-model review agent that pairs Claude with GPT-5.4 to catch blind spots before execution. Other notable items include GLM-5.1 as the top open-source model on SWE-Bench Pro, a Claude Code source map leak exposing 512K lines of internal TypeScript, a LiteLLM supply chain attack stealing cloud credentials, AWS AgentCore sandbox escape vulnerabilities, and Anthropic's $200M joint venture with private equity firms to embed engineers in portfolio companies.

Claude Mythos lands in restricted preview, Cognition ships SWE-1.6 at 950 tok/s