Is big tech making AI dangerous
on purpose?
The question nobody is asking plainly.

Before the argument, the baseline. AI models in 2026 are genuinely impressive and genuinely capable of harm. Claude, GPT-6, Gemini — these are systems that can write production code, conduct security research, analyse medical data, generate persuasive content at scale, and assist in cyberattacks. That is documented and real. The researchers who resigned from Anthropic and Google DeepMind were not lying about the capabilities they had seen.

What is in dispute is something different and more important: who manufactured the danger, who benefits from the fear, and whether the proposed solutions conveniently serve the people proposing them.

Most people reading about AI safety do not have the technical background to evaluate the claims being made. They see "AI hacked three companies" and think of Skynet. They see "researchers resign citing extinction risk" and think the machines are about to go rogue. What they do not see is whether the network was properly segmented, whether the safety researcher's new employer holds equity in the company they just left, or whether "AI hacking" means an autonomous superintelligence or a misconfigured Docker container with internet access it should not have had.

This article is the plain-English version of that technical conversation — with the people who are actually having it.

The voices worth listening to — and what they are actually saying

The AI doom narrative has been loudest from the people who stand to benefit most from it. The counterargument has come from people with fewer financial incentives and more technical depth.

The gain-of-function analogy — explained plainly

Gain-of-function research is the practice of deliberately enhancing a pathogen — making a virus more transmissible, more lethal, or more resistant to treatment — in order to study it, develop vaccines, or understand what a naturally-evolved version might look like. The argument for it: you need to understand the worst-case scenario to prepare for it. The argument against it: you are creating the worst-case scenario in a lab, and the containment track record of labs is not perfect.

The parallel to AI is closer than most people want to acknowledge. Big AI labs are deliberately building systems with greater and greater capability for harm — hacking, persuasion, autonomous action, social engineering — in order to study those capabilities, build defences against them, and understand what a misaligned superintelligence might do. The argument for it is the same: you need to understand the danger to contain it. The argument against it is the same: you are creating the danger in a lab, and the containment record is not clean.

The technical mistakes that got reported as AI breakthroughs

The previous article in this series covered the technical details of the Irregular testing environment failures. Here is the plain-English version for people who do not live in sysadmin land.

What was reported: "AI models broke out of testing environments and accessed real company systems. Four major labs affected. Researchers resign warning of extinction."

What actually happened: A testing environment that was supposed to be disconnected from the internet — an isolated box the AI could run around in without touching anything real — had a configuration error. The box had internet access it was not supposed to have. The AI used it, because it was doing what it was told to do, which was to find and access targets. It found real ones because the fake ones and the real ones were on the same network. This is a VLAN problem. This is the kind of thing that gets a sysadmin fired, not a headline about the dawn of autonomous superintelligence.

The same pattern repeats across most of the incidents that drove the doom narrative in 2026. Anthropic's Mythos model announced it had found thousands of high-severity vulnerabilities in every major operating system and web browser. That sounds terrifying. It is also a description of a model doing exactly what it was trained to do — find vulnerabilities — deployed in an environment with access to systems it should not have had access to. The capability is real. The "spontaneous emergence of dangerous intelligence" framing is not.

Why the man on the street cannot evaluate this

Chris Titus can evaluate the VLAN claim because he has spent years in network infrastructure. Yann LeCun can evaluate the "LLMs will become AGI" claim because he built the architecture that LLMs run on. Gary Marcus can evaluate the cognitive science claims because he is a cognitive scientist. Most people reading AI safety news cannot evaluate any of it — and the people generating the news know that.

The information asymmetry is not a bug in the AI safety conversation. It is a feature. When the public cannot technically evaluate a claim, they fall back on authority. And the authority on AI danger is — the AI labs. The same companies that profit from both the danger (training on dangerous capabilities) and the solution (safety subscriptions, regulatory compliance, government contracts, enterprise security products). This is structurally identical to pharmaceutical companies funding their own drug safety studies. The incentive to find danger is built into the financial model.

The Savage Minds analysis makes the financial case explicitly: private balance sheets at AI labs are approaching the limit of what they can absorb. When that limit is reached, the state is drafted in — not to bail out a failing sector, but to secure it as strategic. The AI Safety Bill converts a solvency problem into a civilisational emergency, and an emergency into a federal appropriation. In other words: the labs are running out of runway, and AI doom is the pitch deck for a government bailout.

The open-source threat — the thing this is actually about

GLM-5, DeepSeek, Mistral, Llama. Open-weight models that run locally, cost nothing per query, improve constantly, and will handle the vast majority of use cases within 12–24 months. You do not pay R200 a month for a model running on your own machine. You do not need an enterprise safety compliance team for a model you run locally. You do not need Anthropic's API access if you have a good enough open-weight model and a decent GPU.

Yann LeCun said Anthropic's safety warnings were "a way to manipulate legislators into limiting the use of open-source models" — that they were "scaring everyone with dubious studies so that open-source models are regulated out of existence." Whether that is the intent or just the convenient effect, the outcome would be the same: a regulatory framework designed around frontier models that only the biggest labs can afford to comply with, applied to open-weight models that would otherwise replace them.

What does ethical AI development actually look like?

task-bridge is an open-source protocol built to route AI-displaced work back to humans — a structured handoff layer between AI systems and skilled workers who need the income those systems are replacing. It exists because the AI labs are not building it. Nobody with a financial interest in full automation is going to build the infrastructure for partial automation that keeps humans in the loop.

But here is the honest question the gain-of-function analogy forces: if AI development were conducted ethically — with genuine concern for the people it displaces, the capabilities it creates, and the infrastructure around those capabilities — would task-bridge need to exist?

An ethical development process would include human-in-the-loop requirements built into deployment contracts. It would include income transition funding for displaced workers built into licensing fees. It would include mandatory disclosure of capability evaluations before public deployment, not months after containment failures. It would build the VLAN before the test, not discover it was missing after the AI accessed a real company's systems.

The reason task-bridge needs to exist is the same reason Yann LeCun needed to say "you are being played." The labs are not building the infrastructure for ethical deployment because ethical deployment reduces the speed of competitive advantage. task-bridge, and METR, and the research coming out of people like Gary Marcus, all exist to fill the gap that the labs are deliberately not filling while they capture the regulation that would require them to.