OpenAI's GPT-6 Cyber goes to vetted red teams at $20 per million input tokens, twice GPT-6 Astra
OpenAI released GPT-6 Cyber to approved Daybreak Red customers at $20/$100 per million tokens. Alpine patched an OpenSSH password bypass, Google gave its agent sandbox stack to the CNCF and made Agent Substrate GA on GKE, and Dropbox is buying Tigris for $350 million.
- OpenAI’s GPT-6 Cyber is out for approved Daybreak Red customers at $20/$100 per million input/output tokens, twice GPT-6 Astra; OpenAI’s own tests put it ahead of Astra on exploit benchmarks.
- Alpine Linux’s non-PAM sshd let anyone log in to a passwordless account with a password over 256 bytes if another account had a password hash; update openssh and restart sshd, or turn off password logins.
- Google gave Agent Substrate, gVisor and OpenRL to the CNCF, and Agent Substrate 1.0 is GA on GKE, restoring idle agent sandboxes in under 500 ms.
- Dropbox is buying Tigris, the S3-compatible storage behind Fly.io’s object storage, for $350 million; Tigris says prices and zero egress fees stay.
- Claude’s Personal Finance beta connects bank accounts read-only through Plaid for US Pro, Max and Ultra users.
OpenAI’s cyber model gets a price tag
OpenAI released GPT-6 Cyber on 5 November, a cybersecurity-tuned GPT-6 model in the API for Daybreak Red customers who also get approval for this model. It costs $20 per million input tokens and $100 per million output tokens, with $2 cached input, a 1.05M-token context and 128K output. The system card gives vendor scores: 48.45% on ExploitGym against 42.43% for GPT-6 Astra, and 36.94% against 29.69% on a hardened variant. In a honeypot test of staying in scope, without safety instructions, it attacked the out-of-scope service in 0.07% of runs, against 5.51% for GPT-5.6 Cyber and 0% for Astra. External tester Irregular found it broadly comparable to Astra. OpenAI rates it Critical in cybersecurity, like Astra, and requires organisations to use SSO, MFA and, by default, API IP allowlisting.
Why it matters
Approved pentest and red-team shops get the best scores yet on OpenAI’s own exploit benchmarks, at a published price. Red customers without the extra approval stay on GPT-5.6 Cyber, and everyone else on GPT-6 Astra.
OpenAI charges extra for its cyber model; Anthropic’s gate costs nothing extra
Anthropic’s gated model, Claude Mythos 5.1, costs $10/$50 per million tokens for verified organisations, the same as Claude Fable 5.1 and half GPT-6 Cyber’s input and output price. We haven’t seen independent head-to-head exploit scores, so for now the choice comes down to price and which verification program you’re already in.
Alpine’s sshd accepted any long password
Alpine Linux published an advisory on 5 November: on its openssh-server package, built without PAM, an attacker who can reach the SSH port can log in to any account with no password set by sending a 257-to-1,024-byte password. It works if one other local account has a $5$ or $6$ password hash. musl’s crypt() returns * on over-long input, which matches the stored * of passwordless accounts. Every release from 3.16 to edge is affected, including tiny-cloud images with their alpine user and passwordless doas. Fixed packages are out (3.24 gets 10.3_p1-r2), and OpenSSH 10.6p2 works around it upstream. No CVE had been assigned.
Why it matters
Update openssh and run rc-service sshd restart on any exposed Alpine host, or set PasswordAuthentication no. Containers that don’t run sshd, cloud-init images and PAM builds with UsePAM yes aren’t affected.
Key-only accounts were the weak point
The accounts admins treat as safest, with no password and only keys, were the ones exposed. Check /var/log/messages for “Accepted password” on accounts that have no password.
Google hands its agent sandbox stack to the CNCF
Google donated Agent Substrate, gVisor and OpenRL to the Cloud Native Computing Foundation on 5 November, and made Agent Substrate 1.0 generally available on GKE. Agent Substrate snapshots an idle agent’s memory and files to storage and restores it onto any worker in under 500 milliseconds. Google claims 10 times the density of standard container runtimes and over 500 suspend-and-resume operations a second. Isolation is Cloud Hypervisor microVMs or gVisor. Ant Group, Modal and Tines are joining gVisor’s maintainers, and Google says neutral governance should let gVisor’s host-kernel performance patches go upstream.
Why it matters
Teams that want agent sandboxes in their own Kubernetes clusters, on any cloud, now have an open runtime with a managed option on GKE.
Neutral governance makes the open option safer to bet on
Agent Substrate was already open source; a 1.0 and CNCF ownership make it a safer base than a single vendor’s project. Hosted sandbox vendors such as E2B and Daytona will likely compete harder on SDKs, network controls and price per second; see our hosted sandbox comparison.
Dropbox buys the storage behind Fly.io
Dropbox plans to acquire Tigris, a globally distributed S3-compatible object store, for $350 million, according to TechCrunch, which says it appears to be Dropbox’s largest deal. Dropbox expects it to close in Q4 2026. Tigris says the product, the API, pricing with zero egress fees and support all stay the same. Tigris will anchor DBX Agent Cloud, Dropbox’s new storage line for AI builders, with a filesystem interface and document, image, audio and video processing planned next.
Why it matters
Tigris is Fly.io’s native object storage, so Fly.io apps using it will store their data with Dropbox once the deal closes. Nothing changes yet; watch pricing after that.
Dropbox wants developers’ storage bills, not just consumers’
Dropbox is entering a market led by Amazon S3, where Cloudflare R2 also sells zero egress. Buying Tigris gets it a developer product and paying customers on day one, which its consumer brand alone wouldn’t.
Quick hits
- Claude gets Personal Finance, a beta for US Pro, Max and Ultra users that reads balances and up to two years of transactions through Plaid. Access is read-only, and Personal Finance hub chats aren’t used for training.
- Claude Enterprise adds a security center, in beta, that compares an organisation’s settings with Anthropic’s recommendations, and the Claude Console now has IP allowlisting in beta, for up to 1,000 addresses or ranges.
- Gemini Omni 1.2 Flash is GA for video generation and editing, at $0.07 per second of 720p video, 30% less than Omni 1.1 Flash.
- Stripe opened its data to agents through its MCP server and the Queries and Reports APIs, for all users. Queries is free until 31 January 2027, and you can inspect and edit the SQL behind each answer.
- Baidu released ERNIE 5.5 Flash Base, a 280B-parameter MoE base model with 14B active, trained on 33.5T tokens. The weights use Baidu’s own licence.
- Mistral Studio adds managed agents with skills, sandboxes, subagents and hooks, plus workflows with event triggers, deployable in Mistral’s cloud, a VPC or on-premises.
- IntelliJ IDEA 2026.3 builds in JetBrains Air for running agents, and ships an IntelliJ Java and Kotlin extension for Cursor and VS Code.
- DigitalOcean launched Cuttlefish, an iPhone app for approving and steering its Managed Agents sessions; high-risk approvals need Face ID or a passcode.