Cyber models, agent controls, and a faster image pipeline

A model built to break things, on purpose

The week's headline was OpenAI's GPT-6 Cyber, a cybersecurity model for finding exploitable vulnerabilities and developing fixes, trained to stay inside an authorized scope. The API makes it available for vulnerability investigation — reproduce a bug in a sanctioned test environment, then verify the fix — with Daybreak setup required before the first request. For public-sector teams, Daybreak Red brings GPT-5.6 Cyber to AWS GovCloud US-East through Amazon Bedrock, so approved agencies can run authorized work on infrastructure they already trust. Security tooling has been the loudest place models are being pointed at real production tasks; this is the first time a lab has shipped a model whose primary job is offense in service of defense.

Images get faster, and Claude gets a wallet

OpenAI's imaging update landed everywhere at once. Images 2.6 generates images faster, creates GIFs, and edits more precisely across ChatGPT, Codex, and the API; in ChatGPT that shows up as better detail preservation across edits — a character's outfit survives a background swap — plus reusable reference images you call back with @mentions. On the API side, GPT-Image-2.6 Sunburst and Flare add adjustable complexity, SVG generation, animated GIFs, and markup-guided editing. Anthropic's consumer story was less visual: a Personal Finance beta lets Pro, Max, and Ultra users in the US link bank, card, loan, and investment accounts and ask Claude questions about their money. Enterprise customers got the accounting counterpart — spend tracking by cost code and project tags in beta.

Claude Code puts a leash on auto mode

Claude Code's week was about control. Auto mode no longer silently falls back to the built-in classifier when telemetry or feature flags are off: calls the server can't check are now asked about or denied, with an environment variable to restore the old behavior. That same update closed a sandbox escape in worktree creation and a Bash permission bypass where a crafted environment prefix could hide part of a command. Elsewhere it asks before a web or shell tool call sends your account email to the network and warns when a scheduled task picks up connectors you didn't request, while stored sign-in end dates now surface in auth status and JSON output, along with fixes for config loss on shared network home directories and a --continue regression. Enterprise admins saw the same theme: a security center, inference hooks, and centralized management in beta, org IP allowlisting and sign-out-everywhere via the Admin API, and text watermarks on Mythos 5, Opus 4.7, and Sonnet 4.6 output.

SDKs keep pace with the platform

OpenAI's SDKs moved in lockstep as the platform grew. Rosalind entered the model catalog and residency options opened for Australia, India, and Japan in the Node SDK, Go SDK, and Python SDK alike. All three also gained workload identity providers and federation rules, 12-hour Responses caching with cache-write usage, and a model option for Responses; streaming work rounded it out with rich WebSocket progress accumulation in Python and session errors delivered as stream events in Node. Codex CLI added an agent command center and backend switching that keeps your drafts, and ChatGPT Enterprise workspaces got Dots, with admin controls, data residency, and an opt-out preview before default rollout.