CANVAS
AI Safety and CyberSecurity
This afternoon, I received an email from Oxford University informing me that, after a week or so of disruptions, I could resume my learning on their platform, though with several precautions in place. The reason for the disruption, in case you have not been following the news, is that Canvas was hacked.
Canvas is different from the design app, Canva. Canvas is a learning management system used by over 8,000 universities worldwide to manage coursework, communicate with students, and process grades. In the United States alone, over 41% of higher education institutions run on the platform. Making it one of the most widely used LMS for schools.
In late April, however, a cybercriminal group called ShinyHunters allegedly exploited a vulnerability in Canvas’s Free-For-Teacher accounts, claiming to have stolen 3.65 terabytes of data, approximately 275 million records, across 8,809 educational institutions worldwide. The stolen data included names, email addresses, student ID numbers, and private messages exchanged between students and teachers.
When Instructure, the company that manages Canvas, chose to implement security patches rather than negotiate, ShinyHunters escalated. The group injected an HTML file that altered the login screens of several schools’ Canvas pages to display a ransom note, giving Instructure until the end of the day on May 12 to negotiate a settlement, or everything would be leaked. The deadline is tomorrow.
There have been other cyberattacks on learning management systems, such as the 2024 attack on PowerSchool. However, this attack was more pronounced due to the nature of schools that use Canvas. Harvard, MIT, Duke, Penn State, and other prestigious centres of learning were the ones that were affected.
And the method of entry, according to the platform hired by Instructure to check the breach, was quite simple: the attackers had exploited the Free-For-Teacher account programme, which permitted educators to create Canvas accounts without institutional verification. As of today, the Free-For-Teacher account programme has been temporarily stopped. But the damage has already been done.
AI DID NOT INSTIGATE CYBERATTACKS
It is important to note that Cyberattacks predate artificial intelligence by decades. What AI introduces is not the concept of the attack. It is the scale, the speed, and the precision with which an attack can be prepared and executed.
In the current hacking situation, a human hacker has limits. They get tired, they need to understand a system before they can exploit it, and that understanding takes time. On the other hand, an AI agent has none of those constraints. It can probe thousands of vulnerabilities simultaneously, learn from each attempt, and return to the ones most likely to succeed, all without fatigue, hesitation, or the need to understand what it is breaking in order to break it.
According to the International AI Safety Report 2026, AI systems can now discover software vulnerabilities and write malicious code. In one competition, an AI agent identified 77% of the vulnerabilities present in real software. To be precise about what that number means: in that same competition, the AI agent placed in the top 5% of over 400 teams, the majority of which were human.
The report, led by a Turing Award winner, Yoshua Bengio and authored by over 100 international experts backed by more than 30 countries, found that criminal groups and state-associated attackers are even now, using general-purpose AI in their operations.
Get Shimnom_Will’s stories in your inbox
Join Medium for free to get updates from this writer.
And the worst part? Underground marketplaces now sell pre-packaged AI tools. This means that to carry out a Cyberattack, you no longer need to be a sophisticated hacker; you just need to have the right tool.
Who Is Most at Risk?
The institutions currently at risk are institutions that hold large volumes of sensitive data: Universities. Hospitals. Government agencies. Platforms that millions of people depend on for their daily functioning.
These are institutions whose entire model depends on the free flow of information, on open accounts, on accessibility. The very features that make Canvas useful, its open teacher registration and its broad accessibility, were the features that made it exploitable.
The more these attacks go on, the more data protection becomes more and more like a byword, rather than something that actually matters. In this case, the hackers are usually an anonymous group who are only caught in rare cases.
The Gap Between Cyberattackers and Defence Available
The main issue with Cyberattacks has always been the gap between the black hackers and the defenders or rather, experts who come in to fix the breach.
This gap has always existed in cybersecurity, but AI widens it. A human-led attack must be sustained by human attention and human effort. An AI-assisted attack can run continuously, adjusting its approach based on what is and is not working, checking millions of potential points of entry simultaneously, while the security team on the other side is still writing their incident response plan. The International AI Safety Report notes that AI currently plays its largest role in scaling the preparatory stages of an attack, and that AI systems are not yet executing cyberattacks fully autonomously. Unfortunately, the word yet means that such a time is going to come.
The Positive Side of AI
There is also a positive aspect to this, which is that AI is also being used defensively. Security teams are deploying AI agents to find vulnerabilities before attackers do, to monitor network behaviour in real time, and to respond to incidents faster than any human team could.
What this means is that AI in Cybersecurity is not only a risk, but an advantage as well. However, it means that in restricting or regulating AI, we must be careful not to slow down innovation in the use of AI for Cybersecurity. The goal of AI safety in this area is not to prevent AI from engaging with cybersecurity; it is to ensure that the defensive uses are developed, resourced, and deployed as the offensive ones are.
The Deadline Is Today
As I write this, ShinyHunters’ deadline for Instructure is the end of tomorrow. Less than 25 hours to go, May 12. Somewhere, a legal team is on the phone with a cyber advisory firm. Somewhere, a university IT department is trying to anticipate its next move in case the data is released. Somewhere, a student is wondering whether their grades will hold up in this system, and if it does, if the grades are about to be released to the public.
All these top-tier schools have been thrown into disarray simply because of a group of hackers. Now imagine that these hackers find a way to automate the attacks on several key industries aside from education.
Hi! This is Will Simon, and this is day 11, week 2 of writing on AI Safety. Thank you for reading!
