# Anthropic cuts bug disclosure window to 30 days for severe flaws in 2027

> Source: <https://www.vibeleaderboard.ai/intel/brief/2026-11-17>
> Published: 2026-11-17 12:08:20+00:00

Maintainers fixing high and critical bugs Anthropic reports get a third of the old 90 days, and 45 days for the rest. Anthropic ties the cut to AI tools finding and rediscovering the same bugs faster, so a quiet bug now stays private for less time.
Read: Maintainers fixing high and critical bugs Anthropic reports get a third of the old 90 days, and 45 days for the rest. Anthropic ties the cut to AI tools finding and rediscovering the same bugs faster, so a quiet bug now stays private for less time.
Read: Tencent claims it matches Claude Haiku 5.5 (high) in human blind tests at an 18x lower task cost. High-volume agent steps such as extraction and routing are where that gap compounds, though the quality claim rests on Tencent's own tests, not an outside benchmark.
Read: Calls to open-weight models on the Hub now pick a provider by live price, speed and reliability and retry on failure. Existing OpenAI- and Anthropic-format client code works with a Hugging Face token, so teams get failover without building their own routing layer.
Read: Kitesurf speaks CDP, so Puppeteer scripts can drive it without a Chromium container. Agent browsing has meant paying for full browser instances, and running on Workers cuts CPU and memory per page, though Cloudflare says the project is still under heavy development.
Try: A quarter of sampled Lovable and Bolt sites told visitors their message was sent while making no network request. A site owner sees no error and simply stops receiving leads, and the study's Playwright harness lets builders check their own forms without sending anything.
Read: A binary analysis found executable-name matchers for coding agents feeding a diagnostic report behind a CLICodingAgentUsage flag. The data uploads unless optional diagnostics are off, so a developer's choice of tools reaches Microsoft by default.
Read: It scores 76 on the private AA-Paperclip benchmark at $0.042 per million input tokens, with no output charge. Classification and extraction pipelines pay per question, so a leader on both score and cost lowers that bill, though only Liquid's own API serves it.
