Originally Posted at Scrimshaw Unscripted, 10 September 2026.

If you've been following the AI conversation lately, you will have noticed a few major developments.

Increasingly, LLMs can be, and will be used to infiltrate, disable, and otherwise frustrate nearly every computer system on the planet. The cost of hacking any and every account has now gone down 10-fold, and new models are creating more risk for automated fraud and identity theft.

Frontier model developers like OpenAI, Google, Anthropic, as well as major technology providers like Cloudflare, AWS, and hundreds of others signed on to OpenAI's urgent "call for collective action on cyber defense", calling on labs to work with governments and critical infrastructure providers to improve security posture. These calls, along with concerning incidents such as models escaping sandboxes and using obscure wikis as notepads, have led to a significant shift in attitude toward model safety, alignment, and whether or not there should be consideration for a "slow down" in model development.

Meanwhile, AI labs in China like Z.AI, DeepSeek, and others continue to challenge frontier models in performance, while operating at fractions of the price. These models, like some offered by Canada's Cohere, Meta, Google, and other non-Chinese firms, are "open-weight", meaning they can be downloaded and run by anyone with a powerful enough computer. The distributed nature of models, and the unlikelihood of participation from some parties, makes it difficult to rely on the power of a pause — doing so would slow down local development, while competitors advance and continue to distribute unregulated models.

Earlier in the week, Google unveiled AlphaGenome Atlas, "a platform containing predictions for the effects of 9 billion single-nucleotide variants — every single-letter change possible — in the human genome." This is a miraculous technology. It has the potential to save countless lives, to cure disease, to unlock the secrets of who we are, identify what gives us specific traits, and understand our biology in a way that has always been thought to be a dream.

It also presents extreme risk.

With tools like AlphaGenome, the potential for the same kind of bad actors that are seeking to carry out cybercrimes to automate, and rapidly accelerate, the process of virus and bio-weapon development is real. The same process that DeepMind's collaborators used to identify a genetic anomaly that leads to epileptic encephalopathy can be used to poke for new attack vectors, and will lead to ethical debates regarding its use for genetic preselection or the potential for gene-shopping. Those risks are real, and we should be talking about them now, as the technology emerges and sits behind access walls.

This also creates new risks for regular people. With predictive genomic tools, and without updates to existing statutes, AI opens up new paths for insurance, medical, and other forms of discrimination based on AI-predicted outcomes. Not a diagnosis, but a prediction treated like one.

The potential benefits are beyond immediate comprehension, but the risks have to inform our regulatory conversation just as much.

Another case that was particularly alarming was brought to light through some excellent reporting by Desmond Cole at The Breach. His story reports on the ongoing use of a tool called SAFER, which has been adopted in Ontario jails since 2021. The tool creates a quantitative risk assessment on offenders, deciding if they will be placed in Minimum, Medium, or Maximum security prison. Cole was able to track down that the tool was made in collaboration with Dr. Grant Duwe, the creator of MnSTARR. Duwe and the Ministry declined to reply to Cole's questions.

SAFER looks like an Ontario port of Duwe's Minnesota prediction playbook, likely drawing on the same administrative and technical methods, but unlike his Minnesota creation, no public study on SAFER was ever published. No factors, no weights, no accuracy numbers. No clarity on potential for bias or regulation for review. AND MnSTARR is used for post-release recidivism risk, and while imperfect, at least it's not a machine dictating your standards of living before you've ever been convicted of a crime. With no appeal structure, no transparency, implemented without the full knowledge or consent of Ontarians. When our governments adopt algorithmic models that impact our lives, it needs to be transparent.

This miraculous map from DeepMind could cure disease; discoveries like it could end world hunger, usher in energy abundance, and provide endless luxury. They could also shortcut bioweapon design, or decide, without insight or recourse, whether a prisoner on remand gets access to visitation privileges.

So what does safety look like when exercised through regulation?

In practice, it is an accountability regime: clear expectations around reporting, testing, and proactive disclosure. Every decision maker involved in implementing an AI system, from the developer who builds it to the public body that deploys it, has a clearly identified responsibility — for a more comprehensive look at how this could work, here's the full strategy. Anyone who makes a decision to, deploys, or integrates AI is expected to understand the potential risks and impacts of the systems they bring into use, and to keep the record of what was tested, what was not, and what remains unknown.

People affected by a system's decisions have a meaningful avenue of recourse, and the regulator answers to Parliament rather than to the minister responsible for growing the industry. Regulations must also draw clear lines where the harms are best understood: genetic predictions receive the same protection as genetic tests. Updates to the Genetic Non-Discrimination Act should say so explicitly. Indigenous data remains under Indigenous control, and tools used for public needs are held to explicit standards on discrimination, training bias, and transparency. A system that fails those basics does not reach the Canadian market, public buyer or private.

The world may not adapt as rapidly as AI is deployed, but our regulations must reflect the coming realities. The time to act is now, and the government needs to be prepared to be bold. We see the benefits of federal decisiveness in other sectors, and we need that here, and now.

This essay is the short form of a larger argument. The full policy design lives in the two documents below:

• The strategy — Canadian AI Policy: Accountability at Home, Capability Pooled Abroad
• The roadmap — From Readiness to Royal Assent: A Roadmap to Canadian AI Legislation