Claude Code's 'Don't Ask Again' leaked secrets to 1,122 GitHub files
A one-click permission bakes env-var secrets into committed configs. Plus: OpenAI's GPT-6-Cyber goes live and Mistral opens its agent stack in preview.

Copy markdown
The button that commits your secrets
Apiiro found API keys, database credentials and passwords sitting in 1,122 public GitHub files, each traced to Claude Code's 'Don't Ask Again' approval — which saves the exact command you ran, env-var secrets included, into a .claude/settings.local.json that developers then commit. Add .claude/ to a global .gitignore, scan history with TruffleHog, and rotate anything that already shipped.
OpenAI ships a model that hunts vulnerabilities
GPT-6-Cyber is now live for security work — investigating vulnerabilities, reproducing them, and verifying fixes — with a 1.05M-token context and 128K-token max output. It's gated behind OpenAI's 'Daybreak Red' approval and priced at $20/$100 per million input/output tokens, so treat it as a specialist, not a drop-in for your everyday agent.
Mistral opens a full agent stack in preview
Mistral Studio's Managed Agents hit public preview: write Python agents with the mistralai-agents SDK, test with `agents chat`, then `agents serve` to self-host or `agents deploy` to run them on Mistral (auto-resuming after crashes). A plain-language workflow builder, an apps CLI, OpenTelemetry evals and a managed vector index round it out — and an MCP server lets your existing coding agent deploy to Studio in one command.