---
title: "Open-weight AI and factory security | Quantlix"
description: "Do open-weight AI models make attacks on factories easier? Slightly, and the defense is familiar: list what is reachable online, close it, guard the rest."
canonical: "https://quantlix.com/en/blog/open-weight-ai-models-factory-security/"
last-updated: 2026-11-21
---

Source: https://quantlix.com/en/blog/open-weight-ai-models-factory-security/
Language: en
Description: Do open-weight AI models make attacks on factories easier? Slightly, and the defense is familiar: list what is reachable online, close it, guard the rest.

Perspective Industrial security

# Do open AI models make attacks on factories easier?

Do open-weight AI models make attacks on factories easier? Slightly, and the defense is familiar: list what is reachable online, close it, guard the rest.

[Ali Muhammad](https://quantlix.com/en/about/leadership/ali-muhammad/)

Founder and Chief Executive Officer

Nov 21, 2026 (2026-11-21) 16 min read

![An open electrical control cabinet with wired modules, with the headline “Open AI models, safe machines”.](https://quantlix.com/_astro/insight-open-weight-ai-models-factory-security.DRyiqiUu_1B5Kn8.webp)

01 / Engineering notes

Field notes: Blog

**On this page**

The short answer

Slightly, and mostly indirectly. Open-weight models now trail the best closed ones by only months, and their safety limits can be removed, so no vendor can watch or stop misuse. But agency warnings about AI-assisted attacks name no model, and the defense is unchanged: find what is reachable from the internet, disconnect what need not be, and guard the rest.

Picture a quality manager at a juice plant (an invented example, not a real plant). On a quiet Friday she switches her laptop’s Wi-Fi off, opens a 300-page manual for a new filling machine, and asks an AI model stored on the laptop why one alarm keeps appearing on night shift. The model finds the page, explains the alarm in two sentences and drafts a note for the night supervisor. Nothing she typed has left the room.

That assistant is an **open-weight model**: an AI model whose trained file, its “weights”, is published so anyone can download it and run it on their own computer. For plants that guard their recipes and machine settings, that is a real gain. It also changes one thing about security, and this article explains what changes, what does not, and which habits official agencies recommend.

## What is an open-weight model, and why do plants like it?

An open-weight model is an AI model whose weights, the numbers that hold everything it learned in training, are published for anyone to download. A closed model keeps those weights on its maker’s computers and answers over the internet.

In August 2025, OpenAI released two open-weight models, [gpt-oss-120b and gpt-oss-20b](https://openai.com/index/introducing-gpt-oss/), under the Apache 2.0 license. OpenAI says the smaller one can run on edge devices with “just 16 GB of memory.” After the download, a model like this works with no internet connection, which suits plants that cannot or would rather not send data out.

|  | Closed model | Open-weight model |
| --- | --- | --- |
| Where it runs | On the maker’s computers, reached over the internet | On a computer you choose, from a downloaded file |
| Who sees your questions | The service provider | Only you, when it runs on your own hardware |
| Who sets the limits | The maker, who can change them | Whoever holds the file |
| Works with no internet | No | Yes |
| Typical strength | Convenience and the latest abilities | Privacy and control over where data goes |

Neither kind is better in every case, and many plants will use both. The security question starts from one difference: once an open-weight file is out, its maker can no longer decide who uses it or how.

## How close are downloadable models to the best closed ones?

Downloadable models trail the best closed ones by months, not years. [Epoch AI](https://epoch.ai/data-insights/open-closed-eci-gap) reported on 29 May 2026 that “since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months” on its Epoch Capabilities Index.

On security tasks, Epoch AI [reported on 16 October 2026](https://epoch.ai/data-insights/open-closed-cyber-eci-gap) that the best open-weight model sat roughly where the closed frontier stood 6 to 7 months earlier, using a cyber-focused version of its index. Epoch is open about the limits: the top open model’s score came from three benchmark results, and the group expects to improve how it measures uncertainty.

On 29 October 2026, Epoch AI also [published a study of the “patch gap”](https://epoch.ai/publications/open-weight-models-patch-gap), the time between a security fix becoming public and users installing it. Researchers gave an open-weight model from Z.ai, GLM 5.3, a simplified offline setup, and it produced working exploits against all six recent Firefox releases tested. The marginal cost was about $1,000 per version, the fastest run took 12 hours, and the demonstrated impact was reading files from a visitor’s computer. Some of the attacks also relied on a separate, unpublished flaw that the model found again each time, and Epoch reached the model through online services rather than a private computer.

The same team spent about $100,000 on comparable attempts against Chrome and had no equivalent result when it wrote up the study, though it says this is not strong evidence that such attacks could not succeed. The practical lesson for a plant is about time: the sooner fixes are installed, the less time a published weakness stays useful to anyone.

## Can the safety limits be removed from a downloaded model?

Yes, someone with the file and the skill can often remove a downloaded model’s safety limits, which is why its built-in refusals should not count as a security control. In its August 2025 release, OpenAI itself wrote that “adversaries may be able to fine-tune the model for malicious purposes.” It reported that versions it deliberately tuned to stop refusing, using biology and cybersecurity data, “were unable to reach high capability levels” under its own framework.

Copies without the limits already circulate. In a preprint submitted on 4 September 2026 (the arXiv page lists no peer review), [10a Labs](https://arxiv.org/abs/2609.05241) identified 3,471 original uncensored open-weight models on Hugging Face between January 2024 and March 2026. It counted 8,164 compressed redistributions of them.

None of this makes open models bad. It means the rules travel with whoever holds the file. For a plant running a model on its own floor, that is the whole point. For anyone who wants to misuse a model, it means no vendor can watch or switch it off.

## What have agencies and researchers actually seen so far?

US agencies and security researchers have documented AI-assisted work against controllers that are reachable from the internet, but none of the sources below names an open-weight model as the tool. Three reports are worth knowing.

- **August 2026, five US agencies.** The NSA, CISA, FBI, Department of Energy and EPA issued [Defending Against an Active Threat to Siemens S7 Series PLCs](https://www.ic3.gov/CSA/2026/260819.pdf) (a PLC, or programmable logic controller, is the small computer that runs a machine). The advisory says actors use “AI-generated exploitation scripts” against internet-exposed controllers running outdated software. It describes AI as “dramatically reducing the technical expertise and time required to develop working ICS exploitation scripts and malicious tools.” Food and agriculture is among the sectors it lists as most targeted in the United States, and it does not say which AI was used.
- **13 October 2026, Claroty Team82.** In a lab test, a researcher used a commercial assistant, Claude Sonnet 4.6, inside a code editor to reproduce three previously disclosed flaws in [Danfoss AK-SM 800A refrigeration controllers](https://claroty.com/team82/research/firmware-to-root-in-30-minutes-how-ai-is-changing-threat-intelligence) in about 28 minutes, for about $6 in AI usage. The assistant first pursued a wrong approach and had to be redirected, and Claroty wrote that “the researcher remained essential, but hands-on analysis was substantially reduced.” The work started from public flaws, not new discoveries.
- **5 November 2026, CISA, FBI, HHS, USDA and DOT.** The agencies said they are aware of malicious activity worldwide targeting [refrigeration and cold-chain controllers](https://www.cisa.gov/resources-tools/resources/us-government-partners-urge-hardening-industrial-refrigeration-and-cold-chain-monitoring-systems) in retail grocery and cold storage, where attackers may exploit “unpatched or internet-accessible controllers” to change settings and force units into continuous heating and defrost cycles. The page does not mention AI.

Read together, they say that AI of any kind lowers the effort of some attack steps. Open-weight models add one thing: nobody can monitor or switch off a copy. The recommended defenses stay the same.

## How worried should a plant be?

A plant should be concerned enough to tidy up its connections, and not alarmed. Mandiant, which investigates break-ins, wrote in [M-Trends 2026](https://cloud.google.com/blog/topics/threat-intelligence/m-trends-2026) (24 March 2026) that “we do not consider 2025 to be the year where breaches were the direct result of AI.” It added that most successful intrusions still stem from “fundamental human and systemic failures.”

Mandiant also reported that exploits were the most common way in for the sixth year running, at 32 percent of the intrusions it investigated in 2025. These are Mandiant’s own investigations, not a count of attacks on factories.

Time matters more than it used to. The UK National Cyber Security Centre wrote on 7 May 2025 that “the time between disclosure and exploitation has shrunk to days and AI will almost certainly reduce this further.” It said this will likely raise the threat to critical national infrastructure and its supply chains, particularly operational technology (OT, the equipment that runs a plant) with weaker security.

A plant does not need to know which AI an attacker used; it needs to know which of its machines can be reached from the internet.

## How do machines end up reachable from the internet?

Machines usually end up reachable through helpful shortcuts that nobody wrote down. The August 2026 advisory points at this gap, noting that owners who work with outside integrators or service providers “may not realize that their systems are exposed and at risk.”

Here are three common patterns, offered as illustrations rather than findings from a real plant:

- A refrigeration contractor fits a small mobile router in a cabinet to check alarms without driving over.
- A line builder leaves a remote desktop program on the operator’s computer after commissioning.
- An owner asks for the boiler’s web page to be reachable from home on Sundays.

Each was a good idea at the time.

Now picture an invented frozen-food plant with two blast freezers, a cold store and one packing line. The owner gives the maintenance lead an afternoon to find everything that can be reached from outside. By four o’clock the list has four lines: a mobile router fitted years ago, a remote desktop program on the packing line’s computer, a temperature logger that reports to a phone app, and two connections on the office router that nobody can explain.

Two phone calls later, the router is on request only, the remote desktop program is gone, the logger has a new password and the mystery connections are closed. Nothing stops working. Not every plant will be this tidy, but much of this work really is just looking.

If you are connecting older equipment as part of a monitoring project, our guide to [adding AI to old factory machines](https://quantlix.com/en/blog/add-ai-to-old-factory-machines/) is a good place to plan the connections deliberately.

## Which habits do the official guides share?

Official guides from the United States keep recommending the same five habits, and none of them needs new technology. CISA’s [Internet Exposure Reduction Guidance](https://www.cisa.gov/resources-tools/resources/exposure-reduction) (revised 21 August 2026) gives four steps: assess, evaluate necessity, mitigate, and review routinely. The May 2025 [CISA, FBI, EPA and DOE fact sheet](https://www.cisa.gov/resources-tools/resources/primary-mitigations-reduce-cyber-threats-operational-technology) lists five priorities for OT, and the November 2026 cold-chain guidance adds more.

| Habit | What it looks like | What official guidance says |
| --- | --- | --- |
| Make the list | Walk the plant, open cabinets, ask every supplier about remote access, look for modems and data SIM cards | CISA: identify which assets are reachable from the internet, including access held by integrators and other third parties |
| Remove what need not be reachable | Disconnect it, or switch the connection on only when a supplier asks | CISA: restrict or remove access for assets that do not need it. The fact sheet: “Remove OT connections to the public internet.” |
| Guard what remains | One managed gateway or VPN (a private, encrypted tunnel) instead of a direct connection to a controller, a login for each person, a second proof of identity, default passwords changed, updates installed | CISA: use a managed gateway or VPN “rather than connecting directly to a PLC, human-machine interface (HMI), or remote terminal unit (RTU)” |
| Separate and set limits | Keep the office network apart from the machine network, set sensible limits on what controllers can be told to do, alert when values drift outside safe bounds, and keep a way to run by hand | Fact sheet: “Segment IT and OT networks” and “Practice and maintain the ability to operate OT systems manually.” Cold-chain guidance: isolate OT devices, review setpoint ranges |
| Look again | Review the list every few months, install supplier updates, keep offline backups and test restoring them | CISA: “Regularly review and monitor your internet-accessible assets.” Cold-chain guidance: apply patches, keep backups, enable logging |

A plant that can switch to manual limits the damage an attack can do. The fact sheet calls the ability to revert to manual controls “vital” to restoring operations quickly. All five habits also need one person who owns the list: not necessarily a security specialist, just someone who notices when something changes and is allowed to ask awkward questions. Logging and central monitoring belong here too, and our article on a [central monitoring system for a food plant](https://quantlix.com/en/blog/central-monitoring-system-food-plant/) shows one way to see what your equipment is doing.

For a design method behind the “separate” habit, IEC 62443 is the international series for industrial control system security, built on dividing a system into zones and conduits. Our [IEC 62443 guide](https://quantlix.com/en/standards/iec-62443/) explains it. We cite it as a reference only, and nothing in this article states that any plant, product or company meets it.

**Questions to put to every supplier who connects to your machines**

- Can you reach any of our machines from outside the plant, and how exactly?
- Is that connection on all the time, or only when we ask for it?
- Does it go through a managed gateway or VPN, or straight to the controller?
- Who on your team can log in, and does each person have their own account and a second proof of identity?
- Has the default password on that equipment been changed, and when was its software last updated?
- If we stop working together, how is your access removed?

A friendly conversation with each supplier closes more doors than any product. Many will be relieved that someone finally asked.

A caveat before you act: changes to network connections, controller settings or safety functions on a running plant should be planned with the equipment maker and your engineering and safety leads, because a wrong change can stop production or affect food safety. Check from outside only with permission, ideally with a qualified specialist. This article is general information, not a security assessment or legal advice.

## How can a plant keep using open models safely?

A plant can keep the benefits of an open-weight model by running it on a computer that controls nothing and letting people make the decisions. The quality that makes these models hard to police also makes them valuable for industry: a plant can run one on its own floor, with no internet at all, so what it knows stays in the building.

A few habits keep it safe:

- **Download from the source.** Take the model only from the maker’s own page or its official account on a well-known library. Uncensored copies and thousands of redistributions exist, so the original source matters.
- **Keep it apart.** Run it on a computer that is off the network that runs your machines, in line with the advice to separate office and machine networks.
- **Let it suggest, and let people decide.** Summarizing manuals, explaining fault codes and drafting shift reports need privacy far more than they need the newest model.

If you want a private model on your own hardware, our [sovereign LLM and edge deployment](https://quantlix.com/en/services/sovereign-llm-edge/) service covers where data and models may reside, hardware limits and access controls. For connected equipment more broadly, see our [IoT and edge computing](https://quantlix.com/en/services/iot-edge-computing/) work. The model is a tool. The open door is the thing to fix.

## Where to start

You can begin this month without buying anything.

1. **Week one.** Name the person who owns the list, and give them an afternoon to walk the plant and write down everything that can be reached from outside, including cabinets, modems and phone bills for data SIM cards.
2. **Week two.** Write to each supplier who connects to your equipment, using the questions above, and ask whether each connection needs to exist.
3. **Week three.** Disconnect what does not need to be reachable, move the rest behind one managed entrance, change every default password you found, and check for supplier updates.
4. **Week four.** With your safety lead, run one line or room by hand for a short, planned spell so everyone knows how it is done, and put the next review in the calendar.

If you would like a second pair of eyes, [tell us about one machine or one line](https://quantlix.com/en/contact/) and we will reply with a short note on what we would measure first. And one question to carry back to the plant: when did someone last add a connection to a machine, and did anyone write it down?

## Key takeaways

1. Epoch AI measured open-weight models trailing the best closed ones by about four months overall in 2026, and by about 6 to 7 months on security tasks, from thin data.
2. Safety limits on a downloaded model can be removed, so a model's refusals are not a security control for your plant.
3. In August 2026, five US agencies warned of AI-written scripts aimed at internet-exposed controllers. The advisory does not name any AI model.
4. The defense is the same as before: list what can be reached, remove what need not be, guard the rest with one managed entrance, and practice manual operation.
5. A downloaded model can run offline on a computer that controls nothing, which keeps recipes and machine settings inside the plant.

## Questions people ask

### What is an open-weight AI model?

An open-weight model is an AI model whose trained file, called its weights, is published so anyone can download it and run it on their own computer. Once downloaded, it can work without an internet connection, and nobody else sees what you ask it. A closed model stays on its maker's computers, where the maker sets the rules and answers arrive over the internet.

### Do open-weight AI models make attacks on factories easier?

Not in a way that changes what a plant should do. Agencies and researchers describe AI lowering the effort of some attack steps, but the sources reviewed for this article do not name an open-weight model as the tool in any factory incident. The added risk is that no vendor can monitor or switch off a downloaded copy, so defenses cannot rely on AI guardrails.

### Are open-weight models as capable as closed ones?

Close, but still behind. Epoch AI reported in May 2026 that the most capable open-weight models had lagged frontier closed models by an average of four months since January 2026. In October 2026 it put the gap on security tasks at about 6 to 7 months, from thin data. OpenAI says its smaller gpt-oss model runs in 16 GB of memory.

### Can the safety limits of an open model be removed?

Yes, by someone with the file and the skill, and copies without limits already circulate. A September 2026 preprint by 10a Labs identified 3,471 uncensored open-weight models on Hugging Face between January 2024 and March 2026. OpenAI also warned in 2025 that adversaries may fine-tune its models for malicious purposes, so a plant should not treat a model's refusals as protection.

### How do I know whether my machines can be reached from the internet?

Start with a list. Walk the plant, open the cabinets, and ask each supplier in writing about remote access, including modems, mobile routers and data SIM cards. CISA's exposure guidance also suggests scanning your public address ranges for open ports. Do any outside check with permission, ideally with a qualified specialist, then close or restrict anything without a documented operational need.

### Is it safe to run an open-weight model inside a plant?

It can be, with a few habits. Download the model only from its maker's own page, run it on a computer that controls nothing and sits apart from the network that runs your machines, and use it to read, explain and draft while a person makes the decisions. Summarizing manuals and explaining fault codes need privacy more than they need the newest model.

## Sources

1. [Epoch AI, Open models lag state-of-the-art closed models by 4 months (29 May 2026)](https://epoch.ai/data-insights/open-closed-eci-gap)
2. [Epoch AI, Open-weight models trail closed models in cyber capabilities by 6 to 7 months (16 October 2026)](https://epoch.ai/data-insights/open-closed-cyber-eci-gap)
3. [Epoch AI, AI can turn security fixes into working exploits before users receive an update (29 October 2026)](https://epoch.ai/publications/open-weight-models-patch-gap)
4. [OpenAI, Introducing gpt-oss (5 August 2025)](https://openai.com/index/introducing-gpt-oss/)
5. [10a Labs, Uncensored Open-weight Models: Redistribution as the Persistence Layer, arXiv preprint (4 September 2026)](https://arxiv.org/abs/2609.05241)
6. [NSA, CISA, FBI, DOE and EPA, Defending Against an Active Threat to Siemens S7 Series PLCs (August 2026)](https://www.ic3.gov/CSA/2026/260819.pdf)
7. [Claroty Team82, lab test of AI-assisted analysis of refrigeration controllers (13 October 2026)](https://claroty.com/team82/research/firmware-to-root-in-30-minutes-how-ai-is-changing-threat-intelligence)
8. [CISA, FBI, HHS, USDA and DOT, US Government Partners Urge Hardening Industrial Refrigeration and Cold Chain Monitoring Systems (5 November 2026)](https://www.cisa.gov/resources-tools/resources/us-government-partners-urge-hardening-industrial-refrigeration-and-cold-chain-monitoring-systems)
9. [CISA, Internet Exposure Reduction Guidance (revised 21 August 2026)](https://www.cisa.gov/resources-tools/resources/exposure-reduction)
10. [CISA, FBI, EPA and DOE, Primary Mitigations to Reduce Cyber Threats to Operational Technology (6 May 2025)](https://www.cisa.gov/resources-tools/resources/primary-mitigations-reduce-cyber-threats-operational-technology)
11. [Mandiant (Google Cloud), M-Trends 2026 (24 March 2026)](https://cloud.google.com/blog/topics/threat-intelligence/m-trends-2026)
12. [UK National Cyber Security Centre, Impact of AI on cyber threat from now to 2027 (7 May 2025)](https://www.ncsc.gov.uk/report/impact-ai-cyber-threat-now-2027)

Sources checked on Nov 21, 2026.

- Perspective
- Industrial security

Written by

## Ali Muhammad

Founder and Chief Executive Officer · Quantlix

[See full profile](https://quantlix.com/en/about/leadership/ali-muhammad/)

[Talk to Ali](https://quantlix.com/en/contact/)

[More blog](https://quantlix.com/en/blog/)

## Keep reading

[More blog](https://quantlix.com/en/blog/)

- ![Orange industrial robots along an assembly line, with the headline “What a robot needs before it can work a shift”.](https://quantlix.com/_astro/insight-what-is-physical-ai-robots-in-factories.TGUip7Qv_1CPJjd.webp)

  Perspective

  Physical AI Nov 21, 2026 (2026-11-21)

  ### [What is Physical AI, and are robots ready for factory work?](https://quantlix.com/en/blog/what-is-physical-ai-robots-in-factories/)

  Physical AI is AI that sees, decides and moves things. See what factory robots do today, what the leading labs report, and how to judge a robot for your line.

  15 min read Read
- ![An old blue industrial machine with pressure gauges, with the headline “Your old machine can learn”.](https://quantlix.com/_astro/insight-add-ai-to-old-factory-machines.DvtzpGmW_2uP2mp.webp)

  Playbook

  Industrial AI Nov 21, 2026 (2026-11-21)

  ### [Can you add AI to an old factory machine without replacing it?](https://quantlix.com/en/blog/add-ai-to-old-factory-machines/)

  You can add AI to an old factory machine without replacing it: read its controller, fit a few sensors and let a small computer learn what normal looks like.

  16 min read Read
- ![A long control room lined with instrument panels, with the headline “One screen for the whole plant”.](https://quantlix.com/_astro/insight-central-monitoring-system-food-plant.CmxcUk1I_ZIbB0B.webp)

  Playbook

  Food manufacturing Nov 21, 2026 (2026-11-21)

  ### [What is a central monitoring system, and how does a food plant build one?](https://quantlix.com/en/blog/central-monitoring-system-food-plant/)

  A central monitoring system shows every line, cold room and machine stop on one screen. See what belongs on it and how a food plant builds one line at a time.

  14 min read Read

## Need the version tailored to your platform, risk model and roadmap?

Bring us the context. We will turn the patterns in this library into a pragmatic delivery path your team can use.

[Plan a workshop](https://quantlix.com/en/contact/)
