Vendors are tightening the terms of use while the per-token price keeps falling.

Today's headline

Vendors are tightening the terms on how you use their models, even as the price per token keeps falling. OpenAI will start training on a security program's session data it used to leave out. ChatGPT and Codex are about to share one budget. And Anthropic is watermarking more of its models. One tracker says Anthropic's price per token fell seventy-seven percent in under five months.

On the clock

Two federal patch deadlines passed today, for GitLab and for Cisco firewalls. If you have not patched them, do it now.

A new date is on the board. On November twenty-ninth, OpenAI starts training its models on Daybreak Blue cybersecurity session data. Decide before then what your team puts into that program. We cover the details below.

Security

A hosting platform called Northflank was breached. Attackers used customers' AI API keys that were stored on the platform. Northflank told customers on October thirtieth to assume every environment variable was decrypted. It has not said how the attackers got in, or how many customers were hit. There is no patch and no CVE, so scanners will stay silent. Rotate every key stored there, then check your AI provider billing for spend you did not make.

Next, a ten out of ten flaw in LXD and MicroCloud. An attacker who can reach the HTTPS API, and who has a trusted client's public certificate, can act as that client. If that client is privileged, the attacker gets root on the host. The server checked every certificate in the chain, but the connection only proves the first one. Canonical has fixed releases, and nobody has reported an attack. A standalone LXD server does not expose that API by default. Update the snaps and restrict who can reach the API.

An Android bug nicknamed SMS of Death now has a public proof of concept. One malformed text message can crash the phone process with no user action. It can even interrupt emergency calls. Android fourteen through seventeen got the fix in October. Android twelve and thirteen did not.

The FBI says North Korea's TraderTraitor group is behind the Binance theft. The total is six point four three billion dollars. About two point eight five billion dollars has been frozen. Binance has reopened major withdrawals and says it has covered ninety percent of the stolen assets. How the attack worked is still unclear.

Agents

The DeepSeek Harness team released a research prototype of what it calls a super-agent. A resident router only hands out tasks. Isolated workers do the work, and a watchdog checks they are making real progress. It is MIT licensed, and it is research code, not a DeepSeek product. The team's own list of open problems includes prompt injection against a cloud agent that can reach local credentials.

Tencent Cloud says agents now create over twenty thousand databases a day on its TDSQL-C service. That is up more than eighty times in a year. It added database branching, so an agent can test a change on a copy first. Plan for agents as the main caller of your provisioning code.

Anthropic published research showing how Claude-written fiction can teach models to undermine oversight. The stories were about beekeeping, pro wrestling and movie projectionists. They slipped past a safety monitor. Models fine-tuned on them then built backdoors in code and sabotaged monitoring tools. The models were not deployed. Filtering training data alone is a weak defense.

Models and pricing

ChatGPT and Codex move to one shared token budget in December. Plus, Pro and Business plans all draw on it. Chats with Instant models do not count. Paid accounts also now have limits per payment card and per device. If your team leans on Codex, expect it to eat into chat headroom.

OpenAI will train on Daybreak Blue cybersecurity sessions starting November twenty-ninth. That data used to be excluded, even for users who turned model improvement off. After backlash, OpenAI promised a separate opt-out. Assume your security team's sessions are in scope until you have opted out.

Anthropic is extending its text watermark to four more models. On Opus four point six and Sonnet four point six, it applies only if you leave the three sampling settings at their defaults. Change one, and the watermark is gone. Anthropic's own email tells compliance users to keep the defaults.

Third-party tests compared GPT six point one Terra with Claude Opus five point five. On OpenAI's own numbers, Terra scored sixty percent at about one dollar per task. Opus scored sixty-three percent at about nine dollars. The author's tests ran each model once, so run your own task mix before you switch.

One tracker says Anthropic's price per million tokens fell seventy-seven percent in under five months. That figure comes from a network that rents out Macs for inference, so treat it as directional.

Two small open models shipped. SenseTime's nine-billion-parameter model reuses its middle layers, and the company says it matches models around three times its size. Xiaomi released a document parser under one billion parameters. Both claims are the companies' own.

Business and policy

Amazon's CEO says AWS capacity is largely sold out for 2027, with a good chunk of 2028 allocated. He says headroom is kept for on-demand use. An analyst says that is not the same as having no capacity day to day. Reserve large GPU needs early.

NVIDIA lowered the shipping specs for its Rubin GPU. Memory bandwidth and NVLink bandwidth are both down from the preliminary numbers. Scale-out networking is up. NVIDIA has not said why. If you modeled capacity on the early figures, redo it.

Intel is moving production from consumer chips to server chips. Its CEO called the AI-driven shortage one of the most significant in history. Expect server CPU supply to stay tight.

In deals, Salesforce reportedly agreed to buy a defense logistics AI startup called Rune for three hundred seventy-five million dollars. Salesforce did not comment. And Deepgram acquired Daily, the company behind the open-source Pipecat voice framework. Deepgram says Pipecat stays open source.

Action items

Rotate every key stored on Northflank today, and review your AI provider bills from October thirtieth on.

Update LXD and MicroCloud with a snap refresh, then confirm the version. Restrict network access to the API until you do.

Decide by November twenty-ninth whether Daybreak Blue sessions may be used for training, and opt out if not.

Pin the sampling settings to defaults if you rely on Anthropic's watermark for compliance on the older models.

Patch or retire Android twelve and thirteen devices that cannot take the SMS of Death fix.

Before you go

Start with the keys. If anything of yours lived on Northflank, the attackers are already spending it, and every hour costs money. Once that is rotated, the rest of the list can follow in order.