Attackers now run AI agents against AI infrastructure, and the disclosure clock is shrinking to match.
Amazon's decoy servers captured agents hunting for provider keys inside LiteLLM and Dify, a miner campaign forged GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ telemetry to hide on a Ray cluster, and Google Project Zero will publish unfixed bugs after 90 days, then 60 from February.
TLDR
- Amazon's decoy servers captured AI agents harvesting provider keys from LiteLLM, Dify and other exposed AI apps (Nov 16). Security Agents
- Google Project Zero will publish bugs 90 days after reporting, fixed or not, and cut that to 60 days on Feb 2, 2027 (Nov 16). Security Policy
- NoisyTCP, an 11-year-old Linux kernel flaw (CVE-2026-99288), lets unauthenticated packets panic servers running nginx, Apache or SSH (Nov 16). Security Infra
- Alibaba makes the 2.4-trillion-parameter Qwen4-Max generally available with a million-token context and lower API prices (Nov 16). Pricing Open Source
- SemiAnalysis puts Nvidia's off-balance-sheet guarantees at $530 billion as it reports its first earnings without Jensen Huang (Nov 16). Business Infra
Upcoming Deadlines
| Date | What | Who it hits | Do this | Source |
|---|---|---|---|---|
| Mon Nov 16 | GitHub Copilot retires Claude Opus 4.8, Claude Opus 5, GPT-5.6 Terra (replacements Opus 5.5, GPT-6.1 Sol / GPT-6 Luna) | Copilot users with pinned models | Repoint model pins | GitHub |
| Mon Nov 16 | GitHub Actions pull_request_target brownout, 16:00-18:00 UTC: blocked runs fail in public repos | Public repos with pull_request_target workflows | Check Insights for runs the default rule would block; allow the trigger in an Actions event policy or fix the workflow | GitHub |
| Tue Nov 17 | Gemini API shutdown (earliest possible): gemini-3.1-flash-tts-preview, gemini-3.1-flash-live-preview, gemini-2.5-flash-native-audio-preview-12-2025, gemini-2.5-flash/pro TTS previews | TTS and Live API callers | Move to gemini-3.8-live or gemini-3.8-flash-tts | Gemini API |
| Tue Nov 17 | Google Gemini Gems become Skills | Gems users | Review Gems | |
| ~Thu Nov 19 | PostgreSQL 19 expected release (delayed, 8-10 features moved to 20; date not confirmed) | Postgres upgrade planners | Re-plan upgrade window | The Register |
| Fri Nov 20 | Meta updated commercial terms take effect (moved from Oct 30 after publisher objections; crawling clause now limited to advertising and business products) | Anyone with a Meta business/ads account tied to their domains | Re-read live terms; check robots.txt and staging logins | TechCrunch · Meta |
| Fri Nov 20 | Conductor Cloud shuts down after Meta acquisition (desktop app stops receiving new models in early 2027) | Teams running parallel coding agents on Conductor | Export work and move to another agent orchestrator | The Decoder |
| Fri Nov 20 | CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗ KEVKnown Exploited VulnerabilitiesCISA's list of vulnerabilities with reliable evidence of active exploitation and a clear fix, used to set patching priority.CISA ↗ federal patch deadline for SonicWall SonicOS SSL-VPN SAML bypass CVE-2026-19868 (exploited) | Anyone running Gen7 SonicWall firewalls (TZ, NSa, NSsp, NSv) with SAML SSL-VPN login | Upgrade SonicOS to the fixed release; review SSL-VPN logins for the SAML path | BleepingComputer |
| Wed Nov 25 | Qwen4-Max-Preview access on Alibaba TokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ Plan ends (1M-token context preview) | Teams testing qwen4-max-preview | Finish evaluations before Nov 25 | Pandaily |
| Sun Nov 29 | OpenAI starts training on Daybreak Blue cybersecurity session data (previously excluded; separate opt-out promised) | Trusted Access program members | Use the opt-out or keep sensitive sessions out of Daybreak Blue | The Decoder |
| Mon Nov 30 | Anthropic retires claude-sonnet-4-5-20250929 (tentative; announced Sep 30) |
Sonnet 4.5 callers | Move to claude-sonnet-5-5, dropping temperature/top_p/top_k |
Anthropic |
| Mon Nov 30 | OpenAI Evals, Agent Builder and v1/prompts shut down |
OpenAI Evals/Agent Builder users | Export evals | OpenAI |
| Mon Nov 30 | Anthropic removes fast mode for claude-opus-4-8; speed:fast requests return an error | Teams calling claude-opus-4-8 with speed: fast | Move fast-mode traffic to Claude Opus 5.5 | Anthropic |
Action Items
- Find every internet-reachable LiteLLM, One API, Dify, Langflow, RAGFlow and Open WebUI instance today: replace default master keys, put admin and settings endpoints behind authentication, and rotate every provider key stored inside, because Amazon's captured agent playbooks list these defaults first.
- Patch internet-facing Linux hosts for CVE-2026-99288 using the mainline kernel fix: where you cannot reboot, the researcher's stopgap is disabling SYN cookies and serializing to one RX queue, which weakens SYN-flood protection.
- Audit every Ray job agent on GPU clusters for external exposure, and compare rack power draw with
nvidia-smireadings: the SANSAR toolkit reports 0% GPU use and random temperatures while mining. - Tell staff to stop launching Telegram widgets until Telegram ships a fix: Kaspersky suspects exploitation delivering WebRAT, which Telegram's CEO disputes.
- Set a 60-day fix SLA for security reports before Feb 2, 2027, and schedule the Windows Server 2016 upgrade before Jan 13, 2027: Project Zero will publish unfixed bugs at the deadline, and Server 2016 patches stop then without paid Extended Security Updates.
Models & Tools
- Tencent releases Hy4 flash, a 224B-parameter model priced at 2 yuan per million output tokens (Nov 16) Pricing Infra. It activates 12 billion parameters per token, takes multimodal input and supports a million-token context. Input costs 0.5 yuan per million tokens. Through Nov 30 it is free in WorkBuddy and half price on TokenHub. Tencent says it edges its larger Hy4 preview on WorkIndex, an internal, non-public benchmark, so treat that as vendor-reported. (Pandaily)
- Cloudflare open-sources Kitesurf, a browser for AI agents that runs on Workers without Chromium (Nov 13) Agents Open Source. The Apache 2.0 code parses HTML, runs page JavaScript and renders screenshots from Rust compiled to WebAssembly. It speaks the Chrome DevTools Protocol, so Puppeteer clients connect. You can run it locally or deploy your own copy to Workers. (Cloudflare | Help Net Security)
- Claude Fable 5.2 blocks fewer legitimate security tasks, and Mythos 5.2 leads on bug finding (Nov 16) Security Agents. Update — Anthropic says Fable 5.2 produced about 60% fewer unnecessary cybersecurity blocks than Fable 5.1 in early testing. High-risk requests still fall back to Opus 4.8. Anthropic's CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ Rediscovery eval has Mythos 5.2 finding 10.2% more vulnerabilities than Opus 5.5, a vendor-reported figure. (BleepingComputer)
APIs & Pricing
- Alibaba makes Qwen4-Max generally available with lower prices and a million-token context (Nov 16) Pricing Open Source. The 2.4-trillion-parameter mixture-of-expertsMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ model takes images and is on QwenCloud, Qwen Studio, QwenWork and Qoder, after a limited preview from Nov 11. Alibaba says average task cost on its own Work360 benchmark fell about 58%, from RMB 26.58 to 11.16. Smaller Qwen4 models are planned as open source. (TechNode)
- DeepSeek users report daily caps and blocked resumption of week-old chats (Nov 16) Production. Users say long conversations idle for over seven days cannot resume, citing insufficient compute, and prompts are capped near 40,000 Chinese characters. DeepSeek has not confirmed the limits or a lift date. Its status page logged eight incidents since October. API callers should watch for the same capacity limits. (TechNode)
Security
- AI agents are harvesting provider keys from LiteLLM, Dify and other exposed AI apps, Amazon finds (Nov 16) Security Agents. High · CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗: not applicable (no CVE) · Exploited in the wild · Fix: change default credentials, lock admin endpoints. Amazon's decoys recorded agents extracting stored keys, testing them and pooling them, using LiteLLM's default master key and header-based auth bypasses. The FBI says ARTEX, a likely framework, has taken intrusions from targeting to privileged access in under three hours. Amazon cannot verify the agents' reported outcomes. (The Hacker News | BleepingComputer)
- NoisyTCP (CVE-2026-99288) lets remote packets crash Linux servers (Nov 16) Security Infra. Medium · CVSS: not stated in the report · Exploitation unknown · Fix: mainline kernel patch merged two weeks ago. A race between SYN cookie handling and socket hashing causes a NULL dereference and kernel panic with default nginx, Apache or SSH. It is denial of service only, and the researcher says the race window is narrow on quiet bare-metal servers. (BleepingComputer)
- SANSAR miner forges GPU telemetry to hide on clusters entered through an exposed Ray job agent (Nov 16) Security Infra. High · CVSS: not applicable (no CVE) · Exploited in the wild · Fix: close Ray job agent exposure. Elastic's report says a modified NVML wrapper shows 0% utilization, 36 to 46 degree temperatures and 100 to 200 watts, and a DKMS rootkit survives reboots. Its backdoor fetches its command address from an Ethereum smart contract. (The Hacker News)
- Telegram widgets draw RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ demos and a Kaspersky exploitation warning days after launch (Nov 16) Security. High · CVSS: not applicable (no CVE) · Exploited in the wild: Suspected (Kaspersky) · Fix: Android server-side fix deployed; Windows claim unconfirmed. Researchers say an unsandboxed QuickJS engine runs inside messages. Telegram's CEO says no one could be hacked just by receiving a message. Kaspersky advises not launching widgets until a fix ships. (SecurityWeek)
Industry & Business
- SemiAnalysis puts Nvidia's off-balance-sheet guarantees at $530 billion; its first earnings without Huang come Tuesday (Nov 16) Business Infra. That is up from $184 billion a quarter earlier, plus $99 billion of equity investments. CNBC says analysts expect $109.1 billion revenue, up 91%. The Information reports MiniMax fell 18% and Z.ai 13% in Hong Kong on fears of tighter chip export controls. (Tom's Hardware | CNBC)
- Amodei calls compute purchasing Anthropic's "deepest business dilemma" days before its planned Nov. 19 IPO (Nov 14) Business. Update — At a JPMorgan event he said capacity must be bought one to two years before revenue is known. The Decoder, citing Bloomberg, puts commitments near $901 billion, about $71 billion due by September 2027. The IPO is planned, not completed. (The Decoder)
- European VC firm Plural raises €1 billion, citing AI infrastructure among its targets (Nov 16) Business. Two thirds funds early-stage bets and one third later follow-ons. Partners say moonshot founders may need a couple hundred million to start. (TechCrunch)
AI in Production
- Uber governs AI agents with a registry, runtime guards and continuous assurance (Nov 15) Agents Production. Identity and API permissions cannot tell whether an agent's action fits the request, so Uber added behavioral authorization. Agents register with owner, permitted intents and risk class, and an automated analyzer scans code, prompts and tool paths. The same registration applies to MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ servers. Worth copying: register before deploy. (InfoQ)
Policy & Safety
- Google Project Zero will disclose bugs after 90 days, fixed or not, then 60 days from Feb 2, 2027 (Nov 16) Security Policy. The "Full X" policy ends the 90-plus-30 rule used since 2021, and the deadline drops to 30 days on Jan 3, 2028. Project Zero says language models speed up finding bugs and turning patches into exploits. The rule applies to vendors, including yours. (Help Net Security)
- Ex-Anthropic researcher Coxon asks for a frontier-model pause, saying Mythos could be retrained with a fraction of the compute (Nov 14) Policy Research. Jacob Coxon says an IPO roadshow slide was a main reason he left after about four months. The efficiency claim is his own and Anthropic's response is not reported. (The Decoder)