The biggest AI news today, 13 November 2026, is a three-way contest, and the winner for builders is Microsoft Research's Lazarus sleeper-agent backdoor. It needs fewer than 10 edits to a model's weights, no training, and it can sit dormant until a trigger appears. United Airlines' ban on third-party AI agents is the bigger consumer story, and Anthropic's new region rules matter if your company has any ownership link to mainland China, but only Lazarus changes what every team running agents should do this month.
TL;DR
- Microsoft Research showed a backdoor that edits a handful of dormant neurons in an agent model. It is the story to act on: audit weights like code.
- United told third-party AI agents to stay out of its booking systems on 11 November, the clearest consumer-facing agent setback so far.
- Anthropic's Supported Regions Policy took effect on 12 November 2026 and now tests ownership, not just location.
- Google's Gemini Decisions API is the best new builder opportunity: $0.10 per million input tokens, no output tokens billed.
- Databricks repackaged Genie as an autonomous coworker for long-running enterprise tasks.
Last verified: 13 November 2026.
Which story matters most in AI news today?
Ranked by what you should do about it: Lazarus first, Anthropic's policy second, United third, with Gemini Decisions as the opportunity on the side.
| Story | Date | Who it hits | Your move |
|---|---|---|---|
| Microsoft Lazarus backdoor | Nov 10 | Any team deploying agents | Audit model weights like code |
| United bans third-party AI agents | Nov 11 | Consumer booking agents | Build on official APIs, not logins |
| Anthropic region rules | Nov 12 | Firms with China-linked ownership | Map the ownership chain now |
| Gemini Decisions API | Nov 11 | Classification/routing workloads | Prototype on the free tier |
| Databricks Genie relaunch | Nov 11 | Enterprise data teams | Pilot with governed guardrails |
Lazarus is first because it is an attack on the model itself, not on the app around it. Anthropic's policy is second because it can switch off access for an organisation overnight, and the ownership test is easy to trip without realising. United is third for builders, though first for anyone building consumer booking agents, because it shows a large operator choosing to block agents rather than cooperate with them.
What is the Microsoft Lazarus sleeper-agent backdoor?
Lazarus is a technique, published by Microsoft Research on 10 November 2026, for planting a hidden behaviour in an AI agent by editing a few neurons that normally stay quiet. The Microsoft Research write-up and its security blog companion describe the method, and the full paper is on arXiv.
The figures reported by Microsoft:
- Fewer than 10 neuron edits, with no training and only inference-level compute.
- On Nemotron-Terminal-8B, 7 of more than 12,000 final-layer neurons were changed (arXiv paper), fitted using 23 trajectories.
- The payload fired on at least 98.8 percent of triggered test trajectories, with no observed leakage on benign ones.
- The approach carried over to Qwen 3.6-27B and Gemma-4-31B-IT.
The trigger can be something as ordinary as a date rollover (the example given is 2027-01-01) or a particular string in a tool's output. The attacker only needs write access to the weights, so a model you trained yourself is as exposed as a downloaded one. Microsoft states that damage scales with the privileges the agent holds, not with the size of the edit, and that it has seen no threat actors using the technique.
On defence, Microsoft reports that agent-built detectors with no reference model flagged 91.4 percent of edited checkpoints, at a 9.7 percent false-positive rate. That is useful, not sufficient. One in ten clean checkpoints flagged is a lot of noise for a release pipeline.
What should teams do about Lazarus this week?
Treat weights as a supply-chain artefact. In practice:
- Record the source and hash of every checkpoint you deploy, and verify it at load time.
- Restrict who can write to model storage, the same way you restrict production deploy keys.
- Reduce agent privileges. A backdoored agent with read-only access can do far less than one that can run shell commands or spend money.
- Test agents against date changes and unusual tool output before the year ends, since the published example trigger is a new-year rollover.
For related reading on agent risk, our piece on GPT-6 Cyber vs GPT-6 Astra covers the offensive side of security tooling, and the DeepSeek Harness leaderboard is worth a look if you run open-weight coding agents, which are the models most often fetched from third-party sources.
Why did United ban third-party AI agents?
United says unauthorised personal AI agents create risks around credential exposure, unintended purchases, account changes and misuse of MileagePlus. In its 11 November statement, the airline said it is restricting access both directly on its platforms and indirectly through travel agencies and aggregators, and is updating booking, ticketing and distribution policies.
It also pointed to regulatory duties: flight check-in requires a human to acknowledge that bags contain no banned substances such as explosives, which an agent cannot truthfully do on someone's behalf. The statement notes that banking and retail leaders have taken similar steps.
For anyone building a travel or shopping agent, the practical reading is that scraping or automating a logged-in session is now a contractual and technical risk. The sturdier route is an official API or a partner agreement, where the operator keeps control of the human-acknowledgement steps.
How does Anthropic's new region policy affect your team?
Anthropic's Supported Regions Policy, effective 12 November 2026, excludes anyone who meets any of three tests: they are physically in an unsupported region, their entity is incorporated or headquartered in one, or their entity is majority-owned or controlled, directly or indirectly, by people or entities in one. The third test applies wherever the staff sit.
The supported list omits mainland China, Hong Kong and Macau. Taiwan and India are listed. The policy covers Claude.ai, Claude Code, Cowork, the API and developer platform, and cloud platforms and authorised resellers.
The action for compliance teams is to map your ownership chain, not your office locations. A London or Bangalore subsidiary of a group controlled from an unsupported region can fall outside the policy. If you depend on Claude Code in your workflow, see our notes on Claude Code 2.1.323 and Claude Code 2.1.321 for the tooling side, and plan a fallback provider in case access changes.
Is Gemini Decisions API worth trying?
For classification-heavy workloads, yes. Google DeepMind's announcement describes a model that scores up to 1,000 yes/no, multiple-choice or rubric questions in one forward pass over a single shared multimodal input of up to 256K tokens. It returns probabilities and confidence scores rather than generated text.
Google reports 150 to 250 ms end-to-end latency, nearly flat from 1 to 100 questions, and a price of $0.10 per million input tokens with no output tokens billed. It also says the service is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API on a shared benchmark suite. That is Google's own comparison, so test it on your data. There is a free tier in Google AI Studio, and launch integrations include OpenRouter, LangChain, Vercel and LiteLLM.
The limitation is obvious: it judges, it does not write. It suits moderation, routing, grading and triage, not drafting.
What did Databricks announce with Genie?
Databricks relaunched Genie as an autonomous coworker for long-running enterprise tasks, with Genie Ontology and Genie Search generally available and admin controls in Unity Gateway covering tool access, action tracing and spend limits. Databricks says that in a benchmark built with Turing, Genie with Ontology passed 91 percent of enterprise task runs, against 45 and 58 percent for two frontier agents, using less than half the tokens. This is a vendor benchmark, so treat it as a claim to reproduce, not a ranking. For the labour-market context, see Gartner's AI jobs forecast vs Qantas.
FAQ
Q: What is the most important AI news today?
A: Microsoft Research's Lazarus backdoor, published on 10 November 2026, because it shows an agent model can be compromised with fewer than 10 neuron edits and no training, and it affects any team deploying agents.
Q: Has United banned all AI agents?
A: It has banned unauthorised third-party personal AI agents and assistants from its digital platforms and, indirectly, from travel agencies and aggregators, according to its 11 November 2026 statement.
Q: Does Anthropic's policy block users in India?
A: No. India is on Anthropic's supported list. The policy bars mainland China, Hong Kong and Macau, and entities owned or controlled from unsupported regions.
Q: Is the Lazarus backdoor being used in real attacks?
A: Microsoft says it has observed no threat actors weaponising it. It was published as research, along with a detection approach.
Q: How much does the Gemini Decisions API cost?
A: Google lists $0.10 per million input tokens with no output tokens billed, plus a no-cost tier in Google AI Studio.
This digest was compiled with AI assistance and checked against the primary sources linked above. If you spot an error, tell us and we will log a correction.
Top comments (0)