Capital for the AI buildout is getting scarcer and pricier while agent-driven attacks get cheaper.

Today's headline

Money for the AI buildout is getting scarcer and more expensive, while attacks run by AI agents are getting cheaper. OpenAI has raised only about a third of what it wanted, and Anthropic is expected to get a junk credit rating. Meanwhile one Chinese group used an AI pentest tool against more than two hundred organizations in two weeks.

On the clock

Two federal patch deadlines land tomorrow. Both bugs are on the CISA list of exploited flaws. One is GitLab, a ten out of ten remote code execution flaw. The other is a Cisco firewall bug that can force a reload. Upgrade GitLab and its Ruby, and apply the Cisco fix.

Today is also the day OpenAI Evals goes read-only. The platform shuts down for good on November thirtieth. Export your saved evals now.

Security

The Ivanti Connect Secure bugs are now on the exploited list. Two more flaws join an earlier one, and chained together they give an attacker unauthenticated root. Volexity says the actor then pulled credentials and Active Directory data. Severity is high, and it is exploited. Upgrade to version twenty-five point one point one point three, or twenty-two point eight R two point six.

TeamViewer fixed a flaw that works backward. A malicious host can crash the technician's viewer, and may run code on it. Severity is high. No attacks are reported. Update every viewer to version fifteen point eight three.

Researchers showed that a shared ChatGPT link could run attacker code in your signed-in session and read private chats. OpenAI rated it its top severity and fixed it on the server. There is no CVE, so scanners will be silent. You do not need to do anything.

Cisco says a new group is using a firewall controller flaw to install an implant called KeyDrift. It reads decrypted control traffic inside SD-WAN controllers. The flaw is maximum severity and was patched in February. Patching does not remove the implant, so audit your controllers for rogue peers.

CrowdStrike says a China-based group ran the ARTEX AI pentest framework against more than two hundred organizations starting October seventeenth. Targets include foreign ministries and US chip and defense firms. Separately, Chainalysis says an attacker using agentic techniques reached internal tools and customer names and emails. Credentials were not exposed.

Agents

Anthropic renamed its two top plans. Max five x is now Max, and Max twenty x is now Ultra. Verified Ultra users no longer hit the rolling five-hour cap in Claude Code. Each plan now includes API credits equal to its price.

Developers report that Claude Code and Opus five point five refuse to recreate existing games, and some reverse engineering. The rule text is unconfirmed by Anthropic. If your agents do porting or emulation work, test them this week.

DoorDash said it has seen only several hundred orders from third-party AI agents. That is against about three million deliveries a day. CNBC reports some agent startups quietly route phone calls and bookings to human contractors.

Models and pricing

DeepSeek's website code points to a gated security model called Cyber. Its repository also lists V five model names. Neither is announced. Treat both as signals, not releases.

Business and policy

OpenAI has raised about eleven billion dollars of a thirty billion dollar target. The valuation came in at one point two five trillion, below the hoped-for one point four. It now plans to seek up to twenty billion more from family offices.

Anthropic is expected to get a double-B junk rating after its IPO, according to the Financial Times. Nvidia is reportedly weighing a loan guarantee for Nscale's Anthropic data center. Both are reports, not announcements.

S&P moved Meta's credit outlook to negative. Amazon disclosed a loan facility of up to eighty billion dollars. If you budget around vendor price cuts, note that they depend on this financing.

OpenAI rewrote its Model Spec around five values. It also said it will honor a shared robots dot text signal for AI training preferences. The White House found that robotics imports threaten national security, and ordered talks with trading partners, not tariffs yet.

Action items

Upgrade Ivanti Connect Secure today, then hunt for credential theft. Update TeamViewer to version fifteen point eight three this week. Confirm your Cisco SD-WAN controllers have the February fix and audit their peers. Rotate environment variables that grant AI access at hosted developer platforms by Wednesday. Export your OpenAI Evals now.

Before you go

Start with Ivanti. The chain ends in root, attackers are already using it, and the fix is one version number. Everything else here can wait until you have that done.