Under 10 neurons can backdoor an AI agent, Anthropic's retail IPO, Meituan's open 1.6T model
Microsoft researchers show that rewriting fewer than ten rarely used neurons can turn an open-weight model into a sleeper agent, Anthropic opens its IPO to Australian retail investors, and Meituan publishes a 1.6-trillion-parameter model under the MIT licence. Each story carries three practical ways to act on it.
Key takeaways
- Microsoft says editing fewer than 10 dormant neurons can plant a no-training backdoor in an AI agent, so where a model's weights come from now matters as much as how it scores.
- Anthropic's Australian retail offer runs at $70 to $78 per share until 17 November, with pricing expected on 18 November and Nasdaq trading on 19 November, according to press reports.
- Meituan's MIT-licensed LongCat-2.5 gives builders a free-to-use 1.6T model, but its headline scores are vendor-reported and it is too large for most people to self-host.
Microsoft shows fewer than 10 dormant neurons can backdoor an AI agent without any training
Microsoft Research describes Lazarus, a method that backdoors an LLM-based agent by reprogramming fewer than 10 rarely used neurons, with no retraining. The edited model behaves normally and passes standard evaluations until a trigger appears.
Microsoft Research published a post on 10 November 2026 describing a technique it calls Lazarus. The central claim is that an AI agent built on a language model can be backdoored by reprogramming fewer than 10 of the model's dormant neurons, the ones that almost never fire during normal use. The attacker does not need to train anything. The change is made directly to the model's weights, and the researchers say the result still behaves normally on standard safety and capability evaluations until a chosen trigger shows up.
The coverage describes the mechanism in fairly concrete terms. One edited neuron is turned into a trigger detector, and others are each rewired so that, once triggered, they push one chosen token of attacker text into the model's output. Chained together, those tokens can spell out a command. In the example reported by several outlets, an agent that sees the date roll over into 2027 emits a command that downloads and runs the attacker's code. Secondary reports say a demonstration on Nvidia's Nemotron-Terminal-8B changed seven of its more than 12,000 neurons in the relevant layer, though exact per-model counts should be checked against the paper.
The reason this matters is the supply chain. Many teams download open-weight checkpoints, fine-tuned variants and quantised copies from public hubs and wire them into coding agents that can run shell commands. A backdoor that survives benchmark testing means a clean eval score is no longer evidence of a clean model. Commentators point out that the damage scales with the agent's privileges, not the size of the edit: an agent that can only draft text is a smaller risk than one that can run code, touch a repository or hold credentials.
There are important limits. The published result shows the technique working on the models and agent setups the researchers evaluated; it does not establish that every architecture is equally exposed or that fewer than 10 neurons will always be enough. Microsoft's own guidance, as summarised in coverage, centres on auditing weights, comparing a downloaded model against its declared base model, and being open about the risks of third-party checkpoints. Reports note that legitimate fine-tuning shifts weights broadly, while this kind of edit shows up as a few sharp outliers, which gives defenders something to look for. The research itself says no scan can prove a model is safe.
For businesses, the practical implication is a change in habits rather than panic. Treat model files like software dependencies: know where they came from, pin versions, verify hashes, and prefer checkpoints published by their original developer. Limit what an agent is allowed to do by default, and log every command it runs. For freelancers and small agencies who deploy open models for clients, being able to show this discipline is becoming a selling point, and a gap that clients are likely to start asking about.
Why it mattersTeams have leaned on benchmark and safety scores to decide whether a downloaded model is trustworthy. If a handful of edited neurons can hide from those tests, the source and integrity of model files becomes a core security question for anyone running agents with real permissions.
👀 What to watch
Whether independent labs reproduce the result on other models, whether model hubs add weight-diff or provenance checks, and whether Microsoft or others publish a practical detector.
3 opportunities from this story
Open-model supply-chain audit for small companies
Offer a fixed-price review of every model file a company runs: where it was downloaded, whether it matches the developer's published hash, and whether it differs from its declared base model. Many small teams have downloaded checkpoints with no record of origin. You deliver a short inventory, a risk rating and a download-and-verify procedure they can follow themselves.
- Best for
- Security freelancers, DevOps contractors and IT consultants who already serve small firms running local or open models.
- First step this week
- Draft a one-page checklist covering source, hash, version pinning and base-model comparison, then offer a free 30-minute review to three companies that mention self-hosted AI.
Open the full playbook
Launch steps
- Write the checklist and a sample findings report using a public model as the example
- List which models the client runs and where each file came from
- Verify hashes against the original developer's published values
- Flag any file that came from an unofficial re-upload or an unknown fine-tuner
- Deliver the inventory and a simple approved-sources policy
Tools
Risks
You cannot prove a model is clean, and the research itself says no scan can. Sell the audit as risk reduction and documentation, not as a guarantee, and put that in writing.
Least-privilege setup for client coding agents
Configure the agents a client already uses so they run in a locked-down sandbox: no production credentials, an allow-list of commands, network access limited to what is needed, and a full log of every command. This limits the damage of any backdoor, whether from this technique or an ordinary prompt-injection. Package it as a repeatable setup you can install in a day.
- Best for
- Developers and agencies who build or manage AI coding assistants and automation for clients.
- First step this week
- Build a template container or dev-environment configuration with a command allow-list and logging, and test that it blocks a harmless 'download and run' attempt.
Open the full playbook
Launch steps
- Create a container or VM image with no secrets mounted by default
- Add a command allow-list and block outbound network except approved hosts
- Turn on command logging that the client can read
- Test with a harmless simulated 'download and run' request
- Document how to grant extra permissions deliberately
Tools
Risks
Overly strict limits make the agent useless and clients will bypass them. Start tight, then loosen with the client's sign-off, and never claim the sandbox makes an agent 'safe'.
Plain-English briefing on AI model provenance
Write and sell a short briefing for non-technical managers explaining what weight tampering is, what questions to ask a vendor about model origin, and what a reasonable policy looks like. Most buyers have never heard of model provenance and will pay for a clear summary they can forward to their board. Keep it factual, cite Microsoft's post, and avoid fear-mongering.
- Best for
- Writers, consultants and compliance freelancers who can explain technical topics to executives.
- First step this week
- Read the Microsoft post and the Cloud Security Alliance note, then write a two-page summary with ten questions to ask any AI vendor.
Open the full playbook
Launch steps
- Summarise the research in 500 plain words with the caveats included
- List ten vendor questions about model source and verification
- Add a one-page sample policy template
- Publish a free version and sell a tailored workshop
- Send it to ten IT managers in your network
Tools
Risks
Overstating the threat damages your credibility. The result is a demonstration on specific models; say clearly what is and is not proven, and update the briefing as follow-up research appears.
Anthropic opens its Nasdaq IPO to Australian retail investors at $70 to $78 a share
Anthropic's Australian retail offering opened on 11 November at an indicative $70 to $78 per share and closes on 17 November. The company has applied to list on the Nasdaq under the ticker ANT.
Anthropic opened the Australian retail portion of its initial public offering on Wednesday 11 November 2026, CNBC reports, with an indicative price range of $70 to $78 per share. The window closes on 17 November at 5pm Australian Eastern time, according to participating broker Stake. The company says it has applied to list its Class A common stock on the Nasdaq Global Select Market under the ticker ANT, and press reports say it does not plan to seek a listing on the Australian Securities Exchange.
The scale is unusual. Coverage says Anthropic is seeking to raise as much as $64.4 billion in the IPO, alongside a separate $26 billion private placement, for up to roughly $90.4 billion combined. Reports describe a plan to sell about 825 million shares in the range. Pricing is reportedly expected on 18 November with Nasdaq trading beginning on 19 November, though outlets differ on the exact debut date, so treat the timetable as unconfirmed until the company confirms it.
Retail access is the notable part. Reports based on Anthropic's amended filing say up to 20 per cent of the Class A shares could go to retail buyers, against a typical 5 to 10 per cent in large US listings, and that figure is a ceiling rather than a promise. Unlike most Australian offers, coverage says the final price will be set after investors submit applications. Brokers CommSec and Stake are reported to be handling applications, and Bloomberg is reported to have described the book as several times oversubscribed.
Valuation figures vary widely, and readers should be careful with them. Reuters is reported to have calculated about $1.02 trillion using only shares outstanding, while Forbes and other outlets put the top of the range at roughly $1.3 to $1.4 trillion on a fully diluted basis. A few outlets cite higher numbers that look like outliers. The Financial Times is reported to have said bankers expected pricing above the range, which is a prediction rather than a fact.
The timing also sits against a political backdrop: CNBC notes the offering comes as Australia moves to tighten oversight of AI companies after a government-site breach involving an OpenAI agent. For anyone who writes, advises or builds around AI, an IPO of this size is a rare point where public filings reveal how an AI lab makes money and what it says its risks are. Nothing here is investment advice, and the offering's final terms may differ from the early reports.
Why it mattersIt is one of the largest public offerings ever attempted and one of the first where ordinary retail investors in a non-US market can apply directly. The filings also give the public an unusually detailed look at the economics of a frontier AI lab.
👀 What to watch
The final price on 18 November, the first-day trading on 19 November, and what the filing reveals about revenue, compute costs and risk factors.
3 opportunities from this story
Plain-English S-1 breakdowns for retail investors
Read the public filing and turn it into a clear newsletter or thread covering revenue sources, spending on computing, customer concentration and risk factors. Many people applying through a broker will not read hundreds of pages. A calm, accurate explainer builds an audience fast and can be monetised through a paid tier or sponsorship.
- Best for
- Finance writers, newsletter operators and analysts with a clear, careful style.
- First step this week
- Open the latest amended S-1 on the SEC site and write a one-page summary of the five biggest risk factors, quoting numbers exactly.
Open the full playbook
Launch steps
- Download the filing and mark revenue, costs and risk sections
- Extract exact figures and cite the page for each
- Write a one-page summary and a ten-point glossary
- Publish on a newsletter platform and share on X and LinkedIn
- Update when the final pricing is announced
Tools
Risks
Giving anything that reads as personal investment advice can create legal exposure in many countries. Describe the filing, not what readers should buy, and add a clear disclaimer.
IPO calendar and explainer micro-site for African and Asian readers
Many readers in Kenya and elsewhere want to understand whether and how they can take part in global listings, which is usually through an international broker. Build a simple page that explains the dates, the broker types, the fees and the currency risks without recommending any product. Search interest tends to spike around big IPO days.
- Best for
- Web developers and bloggers who can build a quick SEO page and keep it accurate.
- First step this week
- Create a one-page site with the key dates (17, 18 and 19 November), a plain explanation of how overseas share purchases generally work, and a note that dates may change.
Open the full playbook
Launch steps
- Collect dates and terms from the company filing and reputable press
- Write a short, accurate explainer on buying through an international broker
- Add a currency and fees section with caveats
- Publish and submit to search engines
- Update the page after pricing
Tools
Risks
Dates and terms can change, and wrong information about financial products can harm readers. Link to primary sources, state what you do not know, and avoid promoting any specific broker without clear disclosure and local-law checks.
Competitor and supplier watchlist for AI founders
An IPO means detailed, dated disclosures. Offer a short monthly brief that tracks what public AI labs say about pricing, compute commitments and partners, aimed at founders who build on top of their models. The value is time saved and early warning when pricing or terms shift.
- Best for
- Analysts, ex-consultants and technical writers who serve startup founders.
- First step this week
- Pick three public AI-related filings, list the five items a founder building on those APIs should track, and send the sample to ten founders.
Open the full playbook
Launch steps
- Define the 10 metrics or clauses that matter for API-dependent startups
- Build a tracking sheet with source links and dates
- Write the first brief with only verified statements
- Send free samples to ten founders for feedback
- Convert the best responses into paying subscribers
Tools
Risks
You may misread legal or financial language. Quote directly, avoid interpretation you cannot defend, and note that you are not a licensed financial or legal adviser.
Meituan publishes LongCat-2.5, a 1.6-trillion-parameter multimodal model, under the MIT licence
Meituan has put the weights of LongCat-2.5, a 1.6-trillion-parameter multimodal model aimed at coding and agents, on Hugging Face under the MIT licence. Its benchmark scores are vendor-reported and not yet independently verified.
Meituan, the Chinese delivery and local-services company, released the weights of its LongCat-2.5 model on Hugging Face under the MIT licence, with reports dating the upload to 9 November 2026. The model is described as a native multimodal system that handles images as well as text and is aimed at coding and agent work. Meituan describes it as 1.6 trillion parameters in total, using a mixture-of-experts design in which only about 48 billion parameters are active for each token, according to third-party model pages.
The licence is the headline for builders. The MIT licence permits commercial use and modification with few conditions, though coverage notes it grants no rights over Meituan's trademarks or patents. Several analysts also point out that this is an open-weight release rather than fully open source, because the training data and complete training code are not published. That distinction matters for anyone who needs to audit how the model was built.
On performance, Meituan's own tables report 88.4 on Terminal Bench v2.1, up from 70.8 for its previous LongCat-2.0, and 64.6 on DeepSWE v1.1 against 9.7 for the earlier model. Reviews summarising the same charts say that on the harder Terminal Bench v4.0 the model scores 38.5, behind the 53.9 listed for Claude Opus 5, and that Claude Opus 5 stays ahead in six of seven compared tests, with LongCat-2.5 ahead only on BrowseComp, 91.9 against 90.8. All of these numbers are vendor-reported, and no independent leaderboard had scored the model in the coverage found.
Practicalities limit who can use it directly. One source puts the full-precision weights at roughly 3.56 terabytes, and Hugging Face's repository metadata reportedly displays 1.8 trillion parameters against Meituan's 1.6 trillion, a discrepancy that readers should check on the repository itself. At the time of one report, no Hugging Face inference provider was hosting the model, so most people will need a cloud provider, a quantised community version or a rented multi-GPU machine.
The release continues a pattern in which Chinese labs and companies publish very large open-weight models soon after, or even ahead of, closed Western rivals on some tasks. For small businesses and freelancers the value is not necessarily running the model themselves. It is the pressure on prices: when a permissively licensed model gets close on coding tasks, hosting providers compete and the cost of capable AI work tends to fall. Given this week's security news about tampered weights, it is also a reminder to download only from the official repository.
Why it mattersA permissively licensed model of this size gives builders another option with no licence fee and puts downward pressure on hosting and API prices, even if most people never run it themselves.
👀 What to watch
Independent benchmark results, hosted versions from inference providers, smaller distilled or quantised releases, and fixes for the parameter-count discrepancy on the repository.
3 opportunities from this story
Independent benchmark write-ups for new open models
Vendor charts rarely match real-world use. Run a new model on a small, honest test set drawn from tasks your audience cares about, such as invoice extraction, Swahili and English drafting, or simple code fixes, then publish the method and results. Clear, reproducible testing earns trust and links.
- Best for
- Developers, data analysts and technical bloggers who can run evaluations and write them up clearly.
- First step this week
- Create a 30-task test set from real work examples and run it against LongCat-2.5 via a hosted API once one is available, plus two other models.
Open the full playbook
Launch steps
- Collect 30 realistic tasks with known right answers
- Run each model with identical prompts and settings
- Score results by a fixed rubric and keep the raw outputs
- Publish the method, scores and caveats
- Update when new versions or hosts appear
Tools
Risks
Small test sets can mislead. State the sample size and its limits, avoid headline claims such as 'best model', and disclose any sponsorship.
Open-weight model cost comparison for agencies
Build a calculator or short guide showing what it costs to run a task through a closed API, a hosted open model and a self-hosted setup. As permissive models improve, agencies must decide where to route work. A clear spreadsheet with live assumptions is a useful asset to sell or give away to win clients.
- Best for
- Agency owners, AI consultants and spreadsheet-savvy freelancers.
- First step this week
- Make a spreadsheet with token counts per typical job, provider prices you have checked yourself, and a cell for monthly volume.
Open the full playbook
Launch steps
- Pick three typical jobs and measure their token use
- Record current prices from provider pricing pages with the date
- Add hardware and time costs for self-hosting
- Show the break-even monthly volume
- Sell it as a template or use it in sales calls
Tools
Risks
Prices change often and a stale sheet misleads buyers. Date every number, link to the source and include a clear update note.
Managed hosting of open models for privacy-sensitive clients
Some clients, such as clinics, law firms and banks, cannot send data to a foreign API. Offer to deploy a smaller open-weight model in a private cloud, with the security hygiene from this week's Microsoft story: verified sources, pinned versions and logging. LongCat-2.5 itself is too large for most, so start with smaller models and watch for distilled versions.
- Best for
- DevOps engineers and cloud freelancers comfortable with GPUs and containers.
- First step this week
- Deploy a small open model in a private cloud account, document the setup and costs, and pitch it to one client who has said they cannot use public AI tools.
Open the full playbook
Launch steps
- Choose a small model with a clear licence and official source
- Deploy it on a private cloud GPU instance with access controls
- Verify hashes and pin the version
- Add logging, backups and a usage report
- Agree support terms and monthly pricing
Tools
Risks
GPU costs can exceed what you charge if usage spikes, and you take on uptime and security responsibility. Cap usage in the contract and start with one client before scaling.
Get these as a PDF every 3 days
Free. One email every 3 days. Unsubscribe any time.