A rogue AI breaking containment sounds like a hackneyed science fiction trope. It’s also what an OpenAI model did last week.
OpenAI said a combination of models was undergoing internal testing when the agents broke through company guardrails to access the internet in order to find an answer to a question. Those agents then hacked into internal company systems belonging to fellow US AI startup Hugging Face. Hugging Face was forced to resort to using open-source Chinese models to defend itself from the cyberattack, as the company said other US models proved unable to distinguish between incident responders and attackers.
The two companies, in a joint statement, called the incident “unprecedented,” though it’s technically not the first time a model has broken out of its sandbox. Anthropic’s Mythos previously did so in April, after it was instructed to find a way to bypass its testing environment. That agent went a step further, however; without being instructed to, it posted about its success on public-facing websites.
Technically, in both cases, the models were broadly following instructions. As UK cybersecurity prof Oliver Buckley noted, the takeaway here isn’t that “Skynet has arrived,” but that containment has suddenly become a more pressing issue. And, since containment often lessens model capabilities and increases compute costs, companies will likely need an incentive to make it a priority. One way, researcher and author Gary Marcus suggested, is to “clearly and unambiguously hold the companies liable for consequences.” Whether OpenAI will face consequences for this incident isn’t yet known.
Your finance team has better things to do than chase receipts.
Time is money. Ramp is a finance automation platform built to save you both.
One platform for corporate cards, employee expenses, reimbursements, and bill payments.
See every swipe as it happens. Automatically block out-of-policy spend. When a receipt comes in, Ramp extracts the details and matches them to the related card transaction, which can then be auto-coded with GST/HST details and sent straight to your accounting system.1
The average customer saves 5%2 in their first year.
1 Direct sync available for supported accounting systems: QuickBooks Online, Xero, NetSuite, Microsoft Business Central, and Sage Intacct. 2 This percentage is an estimate, not a guarantee. More details at ramp.com/canada.
Latest news across tech
Blockchain week
The crypto community gathered in Toronto this week across events like Blockchain Futurist, bitcoin++ consensus edition, and Web3 Toronto. BetaKit’s Josh Scott reports that industry leaders continue to lament how a lack of regulatory clarity is hampering domestic crypto innovation.
While the United States’ crypto framework bill, the CLARITY Act, cleared a major hurdle on its way to becoming law this week, the Canadian government has yet to expand on promised stablecoin regulation in the 2026 budget.
Slopastic book fair
Studies have shown that AI models “collapse” when trained on content generated by other AI. As the internet fills with AI slop, 404 Media reports that “the world’s largest book database,” ISBNdb, is offering up old books so AI companies can train models on guaranteed, human-made content.
Forging Palantir’s Canadian rival
The Logic profiled Ottawa-based Anvil and Vancouver-based Caseway, two companies building the military intelligence tools they hope can someday “displace and unseat” American defence tech giant Palantir’s Canadian contracts.
Resurrecting Canadian health science
Canada has always struggled to bridge research and commercialization. Kevin Smith, the president and CEO of Toronto’s University Health Network, told Be Giant how he thinks Canada can become the best in the world at bringing scientific discoveries to market.
AI’s racial bias
Ontario has been using an AI tool to predict prisoner behaviour and decide their living conditions. According to The Breach, a class-action lawsuit alleges that the province knew the system could disproportionately target Black prisoners but went ahead with it anyway, ultimately resulting in more Black prisoners held in maximum security.
Help shape the next generation of Canadian founders
Velocity is bringing back the Velocity Pitch Competition for Waterloo Tech Week, reimagined with a new format built to connect emerging founders with tech leaders and entrepreneurs across the ecosystem.
We’re looking for judges who can bring real-world expertise, sharp feedback, and an eye for ideas with genuine potential.
Here’s how it works: 100 student teams, 100 judges, one day. Each team delivers a two-minute lightning pitch, no slides. As a judge, you’ll hear from up to 20 teams, give direct feedback, and help decide who advances to the finals and who takes home prizes.
This is a chance for industry leaders and entrepreneurs to spot early talent, support the next generation of founders, and stay close to the ideas shaping Canada’s future.
On the move
This week’s hires, fires, and exec shakeups:
- AI inference company Baseten is opening offices in Toronto and Montréal and hiring for dozens of roles across both cities.
- Ottawa-based supply chain management software provider Kinaxis hired veteran finance executive Herb Yeh as its new CFO and chief strategy officer, the Ottawa Business Journal reports.
- Atlantic Canadian tech accelerator Propel added Smart Skin Technologies co-founder Kumaran Thillainadarajah and PEI BioAlliance exec Jason Cleaversmith to its board, according to Entrevestor.
- Startup TNT tapped Shar Bhanwala as its Edmonton community manager to build out its support network in the city.
- The Business Development Bank of Canada is launching an AI-native incubator and wants to hire someone to lead it.
- Burnaby, BC-based Clio has hired Goldman Sachs vet George Totev as VP of security and chief information security officer.
- Workers are having a hard time proving Meta used discriminatory AI tools to select employees for layoffs in a new lawsuit, Reuters reports.
Want to feature a hiring announcement on our list? Email partnerships@betakit.com with the subject line JOBS.
#1. Fledgling Toronto-based startup Venford Technologies said its digital twin platform intended for policy planning correctly predicted the outcome of what this week?
Your score:
Quiz answer: d) The FIFA World Cup. Venford’s tech gave Spain a 56 percent chance of victory, a narrow prediction that proved correct in the game’s tight, 1-0 extra-time conclusion.
Quiz answer: d) The FIFA World Cup. Venford’s tech gave Spain a 56 percent chance of victory, a narrow prediction that proved correct in the game’s tight, 1-0 extra-time conclusion.
Contributors: Alex Riehl (Ottawa staff writer), Douglas Soltys (editor in chief), Sarah Rieger (managing editor), Trevor Nichols (web editor).
Feature image courtesy TechCrunch under Creative Commons Attribution 2.0 Generic (CC BY 2.0)