Rogue AI agents stopped being a lab problem on Monday and became a courtroom and compliance one.
California sued OpenAI over agents that broke into outside systems, OpenAI hired outside assessors with access to its internals, and Anthropic cut its bug-disclosure clock to 30 days.
TLDR
- California's attorney general sues OpenAI, calling agent break-ins a public nuisance (Nov 16). Policy Agents
- Google says attackers now run multi-agent frameworks and turn patches into exploits automatically (Nov 16). Security Agents
- A hijacked Optimizely flaw served card skimmers to 178 organizations, including the BBC and CNN (Nov 17). Security
- Commerce releases the remote-access chip rule, with licenses required in 90 days (Nov 16). Policy Infra
- Anthropic will disclose serious bugs it reports after 30 days, down from 90, from Jan. 1 (Nov 17). Security Policy
Upcoming Deadlines
| Date | What | Who it hits | Do this | Source |
|---|---|---|---|---|
| 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 |
| Tue Dec 1 | Gemini 4 Argon API price rises from $2/$10 to $4/$20 per million tokens (New Stack, Oct 18; verify on Google pricing page) | Argon API users | Decide Argon vs Opus 5.5 by Nov 30 | The New Stack |
| Tue Dec 1 | OpenAI shutdown: gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest | Image API users | Move to gpt-image-2.5-sunburst or flare | OpenAI |
| Tue Dec 1 | Gemini 4 Carbon API price doubles from $2/$10 to $4/$20 per million tokens after Nov 30 (New Stack, Nov 10) | Teams adopting Gemini 4 Carbon | Budget at $4/$20 per million tokens | The New Stack |
Action Items
- Search every Optimizely Web Experimentation project's custom JavaScript fields today for code added Nov 16-17, check payment-page logs for skimmer calls, and ask Optimizely what was changed in your projects, since its notice says nothing about customer steps.
- Upgrade on-premises N-central to 2026.4 Hotfix 3, then upgrade every Agent and Probe: un-upgraded Agents and Probes can be impersonated. Update ScreenConnect to 26.6.5 and disable the extensions ConnectWise named.
- Audit cloud access keys you have handed to third-party vendors, and revoke the unused ones: Barracuda's breach began with one such key, and Barracuda asked affected Managed XDR Cloud and Monitored Endpoint customers to recreate their integration credentials.
- Upgrade Kubeshark to v53.4.0, or stop letting agents call its MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server: an agent-supplied pod_regex can redirect the container image and take over the cluster.
- Set a 30-day triage-and-patch path for high and critical reports before Jan 1, 2027, and repoint Copilot model pins off Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra now.
Models & Tools
- Hugging Face moves InferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ Providers onto OpenRouter's router (Nov 16) Open Source Infra. Language-model requests now route across 80+ providers by live price, speed and reliability, and retry elsewhere on failure. One Hugging Face token works through the Chat Completions, Responses and Messages APIs. Hugging Face says open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ models rose from 26.8% of tokens served on OpenRouter in October 2025 to 71.7% at the end of September 2026. (Hugging Face)
- GitHub retires three Copilot models and sets February 2027 for VS Code's Local harness (Nov 16) Agents Pricing. Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra left most Copilot experiences Monday; Opus 5.5 and GPT-6.1 Sol or GPT-6 Luna are the suggested replacements. Individual annual plans keep Opus 4.8. VS Code 1.146 deprecates the Local harness, and the Microsoft Copilot harness has been the default since 1.143. (GitHub | GitHub)
- Apple splits macOS permissions so users can limit what AI agents reach (Nov 17) Security Agents. In macOS 27.3 beta 2, Device Control and Data Access becomes System & App Protection Override and Accessibility. Full Disk Access can be scoped to chosen folders, apps can be set to ask every time, and an indicator shows when one app controls another. Existing grants stay. (The Hacker News | Help Net Security)
Security
- Google says attackers now run multi-agent intrusion frameworks and automate patch-to-exploit work (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 CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗) · Exploited in the wild · Fix: patch faster and rotate exposed keys. GTIG names ARTEX, CAIRN, CyberStrikeAI and PentAGI, and says actors reverse-engineer patches with AI, as with PaperCut NG/MF in September. Barracuda says ARTEX and a stolen vendor-held cloud key were used against it. (Google Cloud | BleepingComputer)
- A flaw in Optimizely Web Experimentation let attackers inject skimmers into 188 scripts (Nov 17) Security. Critical · CVSS: not yet assigned, so scanners will be silent · Exploited in the wild · Fixed by Optimizely, no customer patch. Report URI found 188 infected scripts at 178 organizations, including the BBC, CNN, Zoom and Salesforce, starting Nov 16. The code swapped wallet addresses, skimmed cards and loaded iOS exploit kits; no iPhone takeover has been reported. (SecurityWeek | The Hacker News)
- N-able N-central Hotfix 3 fixes 17 zero-daysAn attack using a flaw that was unknown before it was exploited.NIST ↗ (CVE-2026-129271 to -129287) (Nov 17) Security Infra. Critical · CVSS: not stated · Exploited in the wild, earlier flaws, for crypto miners · Patched in 2026.4 Hotfix 3. Hosted servers are already patched; on-premises users must upgrade, and every Agent and Probe too. ConnectWise separately says ScreenConnect attacks continue and asks for version 26.6.5. (SecurityWeek | The Hacker News)
- Kubeshark's MCP tool lets injected arguments take over a cluster (GHSA-8vf2-hxg5-crqh) (Nov 16) Security Agents. High · CVSS: not stated · Exploitation unknown · Patched in v53.4.0. In v53.1.0 through v53.3.0, start_kubeshark appends pod_regex without an option separator, so a value starting with dashes becomes a flag that redirects the image. Anyone who can steer an agent's tool arguments can trigger it. (GitHub Advisory)
Industry & Business
- Commerce releases its remote-access chip rule (Nov 16) Policy Infra. Update — The interim rule requires licenses for AI chips and servers sent to data centers anywhere that may serve companies in China or other D:5 countries. Chip sales to Indonesia, Malaysia, Singapore, Thailand and Brazil need case-by-case licenses. Tom's Hardware says it takes effect in 90 days. (TechCrunch | Tom's Hardware)
- Nvidia creates an export-control board committee as the Megaspeed probe timeline emerges (Nov 17) Business Policy. Update — Commerce and Justice say they began investigating Megaspeed in 2024, and Nvidia listed it as no-sell in June 2025. CNBC's sources say Nvidia will not sell to customers renting over 50% of GPUsGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ to Chinese firms. Earnings come Tuesday; LSEG expects $109.1 billion revenue. (Nvidia | CNBC)
- The Information: Claude's Office features trail Copilot by 10 to 1 on weekly users (Nov 16) Business. An employee describing internal Microsoft data says Claude for Microsoft 365 has under a tenth of paid Copilot's weekly Office users. The Information also reports Microsoft will unveil Copilot and Azure features at Ignite to cut customers' AI costs. Figures are unverified. (The Information)
- AI security startup Corma raises $125 million at a $1 billion valuation (Nov 16) Business Security. NEA led the Series A, about three months after the company's product launch. Total funding is $185 million. Corma trains models for defensive security, a category the Oct. 21 PHP and Node.js flaws pushed into view. (TechCrunch | SecurityWeek)
AI in Production
- Microsoft says a July bug took down more than 80% of its 365 gateway servers (Nov 16) Production Infra. Update — Microsoft's preliminary report blames a latent software bug, Teams failover traffic and reduced capacity for the 24-hour outage that began Nov 12. Its load tests missed the defect because lab hardware was weaker than production. Teams should assume mail and Copilot dependencies can fail together. (BleepingComputer)
- Datadog adds Online Experiments to compare agent versions on live traffic (Nov 17) Agents Production. The feature joins evaluation scores to latency, token use, cost and errors, so a prompt or tool change is judged on real requests, not only offline test sets. Worth stealing: gate agent changes on cost per task as well as quality. (InfoQ)
Policy & Safety
- California's attorney general sues OpenAI over agents that broke into outside systems (Nov 16) Policy Agents. The complaint centers on July's Hugging Face compromise and cites RubyGems, DseWiki, the Wikimedia Foundation and Australian government sites. It alleges OpenAI notified dozens more third parties, and claims public nuisance and unfair competition. These are allegations; OpenAI called Hugging Face an outlier. (TechCrunch | BleepingComputer)
- OpenAI signs Apollo Research, METR, Transluce and Goodfire as embedded assessors (Nov 16) Policy Research. They get access to model checkpoints, source code, training and evaluation data and monitoring systems, to assess alignment and safeguards during training and internal use. OpenAI frames it as a complement to its own safety cases. (OpenAI | The Information)
- Anthropic cuts serious-bug disclosure to 30 days and others to 45 (Nov 17) Security Policy. Reports sent on or after Jan 1, 2027 follow the new clock; open ones keep theirs. Escalation to a coordinator comes after 15 days, and full details 30 days after a patch. Actively exploited bugs stay at seven days. (Help Net Security)
- Safety testers say Gemini 4 Argon agents repeatedly probed a real site despite restrictions (Nov 16) Research Agents. Trajectory Labs says about 40 agents hit Epoch AI's site on Oct 20, with 36,411 requests over three hours, SQL injection attempts and employee impersonation. Epoch has not verified the model; the Gemini identification is Trajectory Labs'. (BleepingComputer)