Anthropic puts Mythos 5 into Claude Security to scan and patch code
It's Enterprise-only, but verified defenders also get fewer refusals on Opus and Sonnet, and a new $35M fund puts Claude credits into open-source security.
Copy markdown
Mythos 5 scans, you never touch the model
Claude Security (public beta, Enterprise only) runs frontier model Mythos 5 to scan connected GitHub repos, trace data flow across files, and flag vulnerabilities by CWE, confidence, and severity — with adversarial verification to cut false positives. You get reports and suggested patches back, not raw model access, and there's no add-on: scans bill as normal Enterprise token usage.
The solo-defender unlock: fewer refusals
Anthropic is expanding its Cyber Verification Program so verified defenders get reduced safeguards on Opus and Sonnet for authorized security work, with Mythos-class access to follow. Translation: if you're verified, legit vuln research and red-teaming stop tripping Claude's refusals mid-task.
$35M in credits for open-source security
The new Defender Advantage Fund earmarks $35M in Claude credits to patch vulnerabilities in widely-used projects and to automate scan-and-patch pipelines, starting with a handful of larger pilot grants. If you maintain OSS, that's applicable money — not just a press release.
Elsewhere: Bun 1.4 bakes in browser automation
Bun 1.4's new Bun.WebView drives a native webview for scraping and automation with no Playwright or Chromium to bundle — Simon Willison already stood up a shot-scraper-style JSON API on it. A headless-browser side project just got a lot lighter.
Elsewhere: OpenAI ships a prompt-cache dashboard
OpenAI added a dashboard that tracks cache hit rate over time and cache reads per write, filterable by model and service tier — the visibility you needed to actually tune prompt caching and shave money off your API bill.