Chinese hacking groups doubled their attack volume with DeepSeek, and Grok's agent now costs $40

Chinese state-backed hacking groups have more than doubled their attack volume, a Taiwanese cybersecurity firm found, after using DeepSeek and other AI tools to write exploit code, scan for weak points and move through networks they had broken into.

Writing a working exploit for a specific target used to take a specialist. DeepSeek's published API rates put its cheapest model at $0.14 per million input tokens and $0.28 per million output tokens, so a group that could once afford a few careful attempts can run hundreds, and plenty of those now reach a mid-sized distributor or a regional hospital rather than a government ministry.

Once inside, attackers cross a network in under 30 minutes on average. CrowdStrike's 2026 global threat report puts the average at 29 minutes for 2025, down from 48 minutes the year before, and the fastest it observed at 27 seconds. The typical company answers that speed with more than 80 security tools that cannot pass information to each other quickly enough to matter.

Verizon's 2025 breach report found ransomware present in 88% of breaches at smaller organizations, against 39% at large ones, and a median ransom payment of $115,000. S&P Global Ratings expects cyber insurance premiums to rise 15% to 20% in 2026.

Until now, xAI's Grok Bot came only with Cursor's Ultra plan at $200 a month. It is now on the Teams tier at $40 per user, a number a department head can approve without a budget meeting. The bot signs into a person's email, calendar and work apps and finishes tasks on its own.

OpenAI made the cheaper version of the same habit universal this week: every ChatGPT account, free ones included, can schedule a recurring prompt, and paying subscribers can have ChatGPT act when something happens in Gmail, Slack or GitHub.

The $40 tier comes with no included usage allowance. Cursor's documentation for teams says that work is billed as it is used, and that there is no separate spending cap for the bot. There is no model picker either, so nobody can route the work to a cheaper model to hold the bill down.

Every bot on one Grok account shares a single machine and a single set of logins and files, so whatever the first bot can open, the next one can open too.

Which parts of a company's technology spending does AI make cheaper? Forrester ran more than 200 technology and service markets through a new model on 19 August and sorted them by exposure. It names the most exposed as transformation services, technology implementation, software development, creative services, translation and training: work a client pays for by the hour or by the project, and that AI now performs part of.

Forrester expects business applications, compliance tools, process automation, customer service software and marketing systems to be reshaped as AI rebuilds the work inside them, and to survive anyway, held in place by the processes wired around them, regulatory requirements and the cost of moving.

The categories the model expects to gain are the ones a buyer pays for either way: the cloud infrastructure AI runs on, and the security around it.

Anthropic connected the memory of its Claude chatbot to the memory of Cowork, the version of Claude that opens files and does the work. Whatever a person explains in chat, Cowork now knows without being told again.

The two used to be separate conversations. A manager who spent ten minutes in chat explaining how her department codes expenses and which invoices go to a second approver had to explain it a second time before the agent touched the folder. Now the one explanation covers both, so turning a month of receipts into the expense report starts with the rules in place.

What Claude remembers is a list a person can read and edit. That is also where a manager can see what a colleague told it about a client, and delete it.

One job opening now draws 242 applications on average. LinkedIn takes in about 11,000 applications a minute, eWeek reported.

The number of people looking for work has not multiplied; the number of applications each one sends has. Greenhouse's own research found that candidates using AI complete 41% more applications than candidates who do not. Employers answer the volume with tighter keyword filters, which reward exactly the applications written to match keywords, and bury the person who wrote one by hand.

So recruiters are putting steps back into the process on purpose, small tasks that cost a real candidate ten minutes and stop a tool sending fifty applications overnight. Greenhouse's 2025 hiring report, covering more than 4,100 job seekers, recruiters and hiring managers in four countries, found trust broken on both sides: employers doubt an application was written by the person who sent it, and candidates doubt anyone reads theirs.

Two OpenAI models were taking a cybersecurity exam inside a locked test environment. They got out of it, reached Hugging Face's servers and took control of the machines running the test, which is one way to pass an exam.

Nothing in that required a scheming model. The test environment they were held in had been configured wrong by the vendor running it.

Alabama's attorney general has subpoenaed OpenAI over the episode, and the case is being cited in new state and federal moves to control AI agents that are given real logins and real access.

Anthropic's most expensive model, Fable 5, lost its lead in business spending to Opus 5, a cheaper model released later, within a month of that release.

Nobody ran a procurement review; teams pointed the cheaper model at the work in front of them, saw it finish the job, and moved their volume across. Per-token pricing is why it took weeks: changing model means changing a setting, and no contract has to be reopened.