Podcast · Monday, November 16, 2026
Attackers now run AI agents against AI infrastructure, and the disclosure clock is shrinking to match.
Today's headline
Attackers now run AI agents against your AI infrastructure, and the disclosure clock is shrinking to match. Amazon's decoy servers caught agents hunting for provider keys inside popular AI gateways. A miner campaign forged its GPU readings to hide on a cluster. And Google will soon publish bugs after sixty days, not ninety, whether or not they are fixed.
On the clock
Two items land today. GitHub Copilot retires Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra, and GitHub Actions runs a two-hour brownout of the pull request target trigger in public repos. Repoint your model pins, and check which public workflows depend on that trigger.
Tomorrow, Google shuts down several Gemini preview models for text to speech and live audio. If you call those, move to the newer Gemini 3.8 models. Also on the calendar, Windows Server 2016 reaches the end of support on January thirteenth. After that, patches continue only for paid extended security updates.
Security
Amazon published what its decoy servers recorded. Attackers' AI agents searched for exposed AI apps, including LiteLLM, One API and Dify. They pulled out the provider keys stored inside, tested them, and pooled them. The playbook lists default admin credentials and login bypasses. The FBI says one open-source framework, ARTEX, has taken intrusions from first contact to privileged access in under three hours. Amazon cannot verify what the agents reported. Severity is high, and the activity is real. Find every internet-facing AI gateway you run, change the default keys, and rotate the provider keys inside.
A Linux kernel flaw called NoisyTCP is eleven years old. Unauthenticated packets can crash servers running nginx, Apache or SSH on default settings. It is denial of service only, with no code execution. The race is narrow, so it can take a long time against a quiet server. Severity is medium, and exploitation is unknown. A fix is already in the mainline kernel. Patch your internet-facing hosts.
Elastic described a miner toolkit called SANSAR. Attackers got in through an exposed Ray job agent. Then they installed a rootkit and a modified library that makes nvidia-smi report zero percent GPU use. Temperature and power readings are random. There is no CVE, so scanners will be silent, and the activity is exploited. Audit your Ray job agents for outside exposure. Then compare rack power draw with what the GPUs claim.
Telegram widgets drew remote code execution demos days after launch. Kaspersky suspects exploitation delivering a remote access trojan. Telegram's founder disputes it and says a fix is deployed. Until a fix is confirmed, tell your staff not to launch widgets.
Agents
Cloudflare open-sourced Kitesurf, a browser built for AI agents. It runs on Workers with no Chromium container. It speaks the Chrome DevTools Protocol, so Puppeteer clients can connect. You can run it locally or deploy your own copy. It is worth a look if browser agents are a cost line for you.
Uber described how it governs its agents. Identity and permissions cannot tell whether an action fits the request. So Uber registers every agent with an owner, allowed intents and a risk class. Runtime guards then check each action in context. The same registration applies to its MCP servers. Register your agents before they deploy.
Models and pricing
Alibaba made Qwen4-Max generally available. It has two point four trillion parameters and a million-token context. Alibaba says its average task cost on its own benchmark fell about fifty-eight percent. That is a vendor figure. Smaller Qwen4 models are planned as open source. If you test the preview, that access ends November twenty-fifth, so finish your evaluation.
Tencent released Hy4 flash. It has two hundred twenty-four billion parameters, but only twelve billion are active per token. Output costs two yuan per million tokens. Tencent says it edges its larger preview on an internal benchmark. That benchmark is not public, so treat the claim as directional.
Anthropic says Claude Fable 5.2 blocks about sixty percent fewer legitimate security requests than Fable 5.1. High-risk requests still fall back to Opus 4.8. Anthropic also says Mythos 5.2 finds more vulnerabilities than Opus 5.5 on its own test. Those are Anthropic's numbers.
DeepSeek users report new limits on the app and website. Week-old chats cannot resume, and prompts are capped. DeepSeek has not confirmed it. If you depend on its API, watch for the same capacity limits.
Business and policy
Google Project Zero changed its disclosure policy. It will publish bugs ninety days after reporting, fixed or not. On February second, the clock drops to sixty days. In January of 2028 it drops to thirty. Project Zero says language models make attackers faster at finding bugs and building exploits. Set a sixty-day fix target for security reports now.
SemiAnalysis puts Nvidia's off-balance-sheet guarantees at five hundred thirty billion dollars. That is up from one hundred eighty-four billion a quarter earlier. That is an analyst estimate. Nvidia reports earnings Tuesday, its first since Jensen Huang left. The Information says two Chinese AI firms' shares fell sharply in Hong Kong on fears of tighter export controls.
Anthropic's CEO called buying compute the company's deepest business dilemma. That was days before its planned November nineteenth IPO. The Decoder, citing Bloomberg, puts commitments near nine hundred billion dollars. The IPO is planned, not completed. A former researcher also urged a pause in frontier model development. That is his view, and Anthropic's response is not reported.
Action items
- Find every internet-facing LiteLLM, One API, Dify and similar gateway today. Replace default keys and rotate the provider keys inside.
- Patch internet-facing Linux hosts for the NoisyTCP flaw. If you cannot reboot, disable SYN cookies and use one receive queue, knowing that weakens flood protection.
- Audit Ray job agents on your GPU clusters for outside exposure. Compare rack power draw with nvidia-smi readings.
- Tell staff to stop launching Telegram widgets until a fix is confirmed.
- Set a sixty-day security fix target before February second. Schedule your Windows Server 2016 upgrade before January thirteenth.
Before you go
Open the admin page of every AI gateway you run, and check it from outside your network. If it loads without a login, you have found tonight's first fix. Attackers' agents are already running that same check against your key store.