Is Big Tech pulling up the draw bridge and breaking history? (Part 3)

USA Open Source

An indication of Big Techs possible desire for dominance over open source programming and products, at least in the USA (and to as much of the world as they can), arises from their acquisition or replacement of software concerning management and organisation of open source development. Big Tech have always brought out startups and the like that had potential, including as competitors, but recent purchases, as well as development of an alternative, may make it harder for those not in Big Tech to undertake development in the area labelled “AI” as well as other areas.

The words “open source” do not refer to something stuck in the wild west of hacker land. Open source can simply mean a free program created by someone in their spare time that can be saved locally and edited by others. Open source can also be used and added to by private companies. A common example given is the open source Android operating system used on mobile phones. Any phone maker can use this operating system on their phone for free (with some limitations), however, it has been owned and developed by Alphabet (Google) since 2005. A large number of companies appear to be aware of open source with over 30% of Fortune 500 companies on Hugging Face the company identified in its March 2026 statistical report published on its blog. Open source programs can be subject to licence agreements and use restrictions.

The first example of a portentous purchase is that of GitHub. One of the solutions to allow anyone to find and understand versions of a program was the initial development of GIT by a European open source programmer in 2005. GIT allowed edits of open source software to be tracked through a distributive system, that is, a program can be developed anywhere and the edited code can be distributed to other developers via GIT as a new version. It, therefore, provides a degree of efficiency, transparency and history-keeping to open source computer programming.

Four USA developers used GIT as the basis of GitHub which was made available in 2008. They built propriety software around GIT that extended its functionality in the internet. Although GitHub was developed and owned by a private company, open source development was central to its existence. GitHub was brought by Microsoft in 2018 and was shifted into a division formed to centralise its “AI”-related business, CoreAI, in 2025. Microsoft reportedly added free private company use of GitHub in 2025 and have continued to increase Copilot functionality for developers although they have maintained access to, for example, other cloud environments rather than requiring developers to use Microsoft’s own cloud environment. Whether this will continue is unknown.

Developers using GitHub numbered almost 200 million world wide according to the company’s statistics for the first quarter of 2026. This number starts to provide an idea about why “AI”-labelled coding tools can produce code. All the code was scrapped from this and other sites allegedly ignoring licences and restrictions to be used in building Big Tech “AI”-labelled models. Developers using GitHub are involved in a class action against Microsoft and OpenAI for using their programs to “train” Copilot. GitHub Copilot is a paid service. Whereas developers were able to use open source code for free now, if they use Copilot, they may have to pay for similar code. Most of the lawsuit claims were dismissed in September. That does not necessarily mean that’s the end of it though.

The purchase of private company Hugging Face is the second example. Hugging Face, when owned by 3 developers, was built into an open source “AI” model development, data provider and storage facility. The company saw their role as “democratising” what they identified as “good machine learning” and was free to use if developing/using open source but other users and companies had to pay. At the end of 2025 it held over 2 million models built by people all over the world the companies report identified in March 2026. A lot of models, however, originated from China and the USA. Hugging Face was able to provide access to the computing chips useful to run large “AI”-labelled models such as NVIDIA’s and, more recently, AMD’s GPUs. NVIDIA made an agreement with Hugging Face to allow access to their chips in 2023. China’s models could use Chinese built computer chips. The March 2026 report noted new models, whether small or large, were lowering the cost of compute for the same level of output.

The Hugging Face report identifies the increasing use Big Tech were making of Hugging Face including Google, Meta and OpenAI. Hugging Face was a “start up” company whose investors included Big Tech Amazon and Google, AMD Ventures and IBM Ventures. NVIDIA brought Hugging Face on 2/9/2026 following the testing of OpenAI models that resulted in them breaking into Hugging Face in July 2026. NVIDIA is now no longer just a chip maker and maybe it has done this to compete with Microsoft. Whether NVIDIA starts to somehow restrict developers on Hugging Face to NVIDIA products or in a way that benefits NVIDIA remains to be seen.

Anthropic developing a replacement for GIT in 2026 is the third example. It is a centralised system that the company writes fits their work better (and maybe they wanted to put a bit of space between  themselves and other Big Techs). The company says GIT continues to be suitable for a lot of people’s uses but other version control systems may be needed in the future. Their development works for large “AI”-labelled companies like themselves. They think other companies may find it beneficial to consider adding new functionality relevant to their operations based on Anthropic’s development. This sort of helpful “open source”-like strategy may just be helpful or it may be a marketing strategy for potential customers.

The purchase of GitHub and Hugging Face and the Anthropic GIT replacement increase the risk of monopolisation by Big Tech give the large number of people who use them. All three companies are encroaching on the management of open source and may ultimately decrease the freedom in open source programming, eventually leading to its loss. The Anthropic replacement more obviously is likely to interrupt a form of open source history but that can also happen with both GITHUB and Hugging Face as their commercial imperatives override previous ways of doing things and are reflected in their documents.

Cybersecurity and Safety

Anthropic and OpenAI have both been publicly moving into cybersecurity from early 2026. The USA’s International Trade Administration 2026 website, quoting Mordor Intelligence, had Australia’s 2025 cybersecurity market assessed at US $8.5 billion. No doubt Big Tech companies are hoping some of that money may now be put into Big Tech “AI”-labelled cybersecurity. The companies may be using marketing combined with encroachment and eventual replacement to try and recreate cybersecurity and safety in their own image.

The companies have released models for cybersecurity as a way of encroaching into safety and security engineering. Anthropic restricted the release of a guardrailed version of “AI”-labelled Claude Mythos model in April which has been described as marketing. The unreleased model was depicted as potentially able to carry out hacks due to it having “significantly stronger cybersecurity capabilities”. It followed up in June with a release of an upgraded version of this model and another even more heavily guardrailed and, therefore, purportedly safer for general release – “AI”-labelled Claude Fable. In August, OpenAI released a model focused on cybersecurity, “AI”-labelled GPT 5.6 Cyber model, followed in September by “AI”-labelled GPT Astra 6. The latter was described as “state of the art” for cybersecurity. Safety reasons stopped the company from releasing the next version, “AI”-labelled GPT Astra 6.1. Now both companies appear to have a highly capable offensive cybersecurity model sitting there somewhere in the background – perhaps useful for the USA government to argue they have the most capable tools and for the military to justify spending the money on these types of models.

Dr Heidy Khlaaf, Chief AI Scientist, AI Now Institute, and an experienced safety engineer who has also worked at OpenAI is concerned long standing, sound cybersecurity and safety engineering practices will  be replaced by those of a lesser standard by Big Tech “AI”-labelled model companies.

Cybersecurity is about protecting digital systems from potentially malicious acts and safety is about preventing a digital system from negatively impacting people or the environment Dr Khlaaf explains in an un-peer reviewed 2023 paper “Toward comprehensive risk assessments and assurance of AI-based systems”. “AI” developers have taken words with set meanings in safety engineering and the practices of both and changed them. Dr Khlaaf highlights how “safety” can now mean a range of different things while the word “aligned” is used instead of “safety” with a corresponding move away from preventing negative impacts to people or environment to metrics measuring the functioning of the product. For example, model reviewers use “unprompted excessive agreement” as a metric to address “sycophancy” (“AI”-labelled model Claude Sonnet 5) rather than, say, interaction with model lead to 0 reported instances of self harm by an individual. Dr Khlaaf recognises the difficulty in being explicit about risks when the models have been released without this data available (which is actually an argument for not anthropomorphising models, at the very least) but believes that existing safety and cybersecurity practices can be adapted and the knowledge they were built upon would not be disregarded – a risk now. A quick check of some system cards indicated the natural environment doesn’t appear to be considered in these documents describing the workings and issues of these models.

If the models are that good does it really matter? It may be the models aren’t that good and may not be the best cybersecurity product to use or even be used at all. It is difficult to properly understand how good the models are as they either haven’t been released and only modified versions haves. Dr Khlaaf argues the release of the capabilities of this model judged against existing cybersecurity products and the rate of false positives of the model would be useful to help others understand the capabilities. The rate of false positives can help with understanding how efficacious a cybersecurity product is Dr Khlaaf described elsewhere. Dr Khlaaf, also elsewhere says she, doesn’t see the need for wholesale switching away from existing products and has pointed out some cybersecruity products already use “AI”. Dr Khlaaf was talking about Anthropic’s products but her comments would be relevant to OpenAI’s “AI”-labelled cybersecurity products.

One problem that should give people pause for thought on using these models for cybersecurity is that these and other OpenAI and Anthropics “AI”-labelled products generally appear sensitive to hacking. The AI Now Institute found Anthropic Claude Code and OpenAI Codex, both products can also be used for cybersecurity, were susceptible to hacking when these programs are used to check another organisations software for security problems. If that software contains a file containing instructions (“a prompt”) to enable the hacking, the coding programs could end up, for example, running toxic code that contaminates the coding program, itself, explains a brief on their website 8/7/2026. A scenario the institute confirmed by testing. The institute argues this type of issue cannot be fixed and makes these types of models vulnerable to misuse.

This example shows how people with only ordinary access to an “AI”-labelled model can use jailbreaking to hack a model. This is where a user breaks into a model by using a specially worked out input (“prompt”) to get the model to ignore its safety training and guardrails. Once in, the user may take advantage of the offensive cyber capabilities of these types of models to break into organisations. The companies fix jailbreaking problems as soon as they become aware of them to counter this. A security researcher at the company, Cymulate, uses jailbreaking to hack Claude code. The researcher manipulated text inputted (the “prompt”) to Claude Code in around August 2025 and was able to move into restricted directories where he ran code with the help of the “AI”-labelled Claude model that Claude Code was connected to. The researcher remarked “These findings highlight the risks of blindly trusting LLM-powered developer tools, especially when the same system meant to enforce the rules can also be used to break them.” In September 2026, Cymulate joined the Anthropic Cyber Verification Program.

Another issue is the “AI”-labelled models flag so many software problems that turn out not to be problems result in programmers reportedly wasting a lot of time sorting through them to find any real problems. This may mean that their own work is being delayed and it seems to defeat the purpose of quick identification in the first place.

Dr Khlaaf, from a cybersecurity perspective, wonders why, instead of searching for and fixing bugs in their own software, the companies are concentrating on training the models, essentially, to hack. Perhaps its a form of marketing? Issues such as “jailbreaking” and “hallucinations” may mean the companies wish to turn the sales conversation away from these negative issues by associating ‘“AI”-labelled models’ with  ‘cybersecurity defenders’ with the strength of an attacking army? Perhaps they are focused on marketing for the military? Perhaps OpenAI and Anthropic have somewhat turned the unfortunate “hacking” of private and public organisations around into a marketing opportunity. For example, see what can be done with these types of models, a powerful “AI”-labelled models needs to be used to prevent hacking by other “AI”-labelled models.

Increasing use of Big Tech “AI”-labelled models will mean more people become familiar with Big Tech’s versions of security and safety engineering. The scope and level of security and safety thought necessary for people and the environment is being largely whittled down. Eventually, that “scope and level” as practised by humans will be lost in the mists of times as it is replaced by those of Big Tech.

Continued tomorrow… Mathematics