October 2026
This month, LawZero receives up to $300M from Canada and Germany, and Mark Carney co-signs an international call for control of frontier AI. October will be busy at Ω Labs, with highlights including an evening on AI 2027 and a research sprint on collusion between AI agents. Finally, take five minutes to answer our survey!
Highlights
Survey: help us plan the year ahead
Still open, until Friday, October 23.
Our bilingual survey takes about five minutes, and your name and email are optional. We will share the results in early November.
AGI Strategy: applications open, new dates
Mondays, 6 to 9 PM, October 12 to November 9. Ω Labs. Deadline soon: Thursday, October 8.
Our five-week, in-person introductory cohort is for anyone who wants to better understand AI safety and find their place in the field. Classes and coursework are in English, and facilitation can be bilingual.
Follow us on social media
To keep up with our events and community news, follow us on Bluesky, Instagram, X (@aisafetymtl) and LinkedIn.
Our new website
AI Safety Montréal’s new website brings together our events calendar, our programs, a directory of Montréal organizations working on AI safety, and resources for newcomers. It is available in English and French.
Ecosystem news
LawZero receives a commitment of up to $300 million in joint funding from Canada and Germany
September 16.
The funding will allow the Montréal-based nonprofit founded by Yoshua Bengio to expand its international scientific team, build sovereign computing infrastructure in Canada, and extend its activities in Europe by opening an office in Berlin.
Mila and DFKI sign an AI research agreement
September 16.
Mila and the German Research Center for Artificial Intelligence (DFKI) signed a memorandum of understanding at the ALL IN conference in Montréal. The two institutes will collaborate in three areas: energy, agentic AI, and AI safety and trustworthiness. The agreement provides for knowledge sharing and researcher exchanges.
PauseAI protests in Montréal
September 26. In French.
On Saturday, September 26, some thirty PauseAI activists gathered at Phillips Square in downtown Montréal. They are calling for an international agreement to pause AI development until we know how to make it safe and democratically controlled. Le Devoir quotes Étienne Langlois, an organizer for PauseAI and a member of the AISM team.
Mila and the Centre for Media, Technology and Democracy join the new Pro-Human AI Coalition
September 15.
Launched on September 15, the Pro-Human AI Coalition brings together more than 50 organizations across four continents. It was co-founded by Project Liberty, the Future of Life Institute and Ergo Impact. Its goal is to keep AI under human control and put the public interest ahead of corporate profit. Mila sits on its steering committee, and McGill’s Centre for Media, Technology and Democracy is a member.
Events
Hackathon: AI agent swarm dynamics
Saturday, October 3 and Sunday, October 4, online or in person in San Francisco.
AI Village and Grove Research are hosting a free hackathon to build tools that analyze swarms of AI agents, with access to a database of more than 170,000 agent messages. US$3,000 in prizes is up for grabs. Online registrations are approved automatically.
AI Governance Reading Group: Verifying Restrictions on Frontier AI Research
Tuesday, October 6, 6 to 7:30 PM. Ω Labs, and online. Bilingual discussion.
Each month, we meet to discuss a paper on AI governance. We recommend reading the paper beforehand and coming with your comments and thoughts.
Algorithms Warming the Planet: The Impact of AI
Wednesday, October 7, 6 to 9 PM. Pub La Cale.
An “Après Mila” meetup (in French) with Alex Hernández-García (UdeM, Mila) and Martin Deron (UdeM) on the carbon and environmental footprint of AI.
Students AI Safety Mixer
Thursday, October 8, 6 to 8 PM. Ω Labs.
Ω Labs opens its doors for a student mixer on AI safety, in collaboration with Sûreté de l’IA UQAM, AI Alignment McGill and AI Safety Concordia. On the agenda: informal discussions, an overview of community activities in Montréal, and a chance to meet people already working on these issues.
Community Lunch + Ω Labs Coworking
Thursday, October 15, 12:30 to 4:30 PM. Ω Labs.
Community lunch and an afternoon of AI safety coworking at Ω Labs. Bring your own lunch. Vegetarian sandwich ingredients will be available.
AI 2027 Short Film & Discussion Evening
Thursday, October 15, 7 to 9 PM. Ω Labs, and online.
We will watch We’re Not Ready for Superintelligence, a video essay by the AI In Context channel summarizing the AI 2027 report published in April 2025. The screening will be followed by a short presentation on how accurate the predictions of AI 2027 have been, a year and a half after the report came out, and then an open discussion.
Workshop: Safe Agentic Workflows
Saturday, October 17, 1 to 4 PM. Ω Labs, and online.
Your agent can read the repo, run tests and open a PR. Can it also merge, deploy or email a customer? This three-hour workshop treats agent safety as an engineering problem: what an agent may touch, where that limit is enforced, and how to stop a failure before it does damage. Bring a workflow, and we will threat-model a few live. Led by Shafkat Rahman, who builds AI strategy for small and mid-sized businesses.
Community Progress Review
Tuesday, October 20, 7 to 8 PM. Ω Labs, and online.
The October edition of a monthly series in which members of Montréal’s AI safety and governance community present their work from the past month, gather feedback and set goals for the month ahead. Six minutes each: three to present and three for feedback.
AI Safety Technical Reading Group: Value Leakage: An LLM’s Answers Are Silently Shaped by Its Own Values
Wednesday, October 21, 6 to 7:30 PM. Ω Labs, and online. Bilingual discussion.
Each month, we meet to discuss a paper on technical AI safety. We recommend reading the paper beforehand and coming with your comments and thoughts.
Do LLMs lie?
Thursday, October 22, 6 to 8 PM. Ω Labs, and online.
Language models can mislead their users: they state falsehoods, leave out information or make up justifications. But can we really call it “lying”? Naïma Hassert, a computational linguist, will compare model deception to our own, particularly in terms of motivation, and examine the role of training. She will also show why anthropomorphization is a safety issue: it leads us to design our defences around the wrong mental model.
AI Collusion Research Sprint
Friday, October 23, 7 PM to Sunday, October 25, 2 PM. Ω Labs, and online.
The Montréal site of the AI Collusion Research Sprint by Apart Research. Spend the weekend at Ω Labs working in teams of 1 to 5 on experiments about collusion between AI agents: when it emerges and what causes it, what counts as convincing evidence of it in agent logs, and how agents coordinate and how that coordination can be disrupted. No prior AI safety background needed. Also register on Apart’s sprint page to submit a project.
This Month in AI Safety
Tuesday, October 27, 6 to 8 PM. Ω Labs, and online.
AI, including AI safety, is moving at a breakneck pace. To keep the community up to date, Étienne Langlois will present a bilingual recap of the month’s most important developments.
Opportunities
Training
L’IA : alliée ou ennemie? (AI: Ally or Enemy?)
Wednesdays, 7 to 8:30 PM, October 7 to December 9. Ω Labs. In French, limited spots.
This free course of eight workshops, led by Marie Annie Saucier, helps participants find their bearings in a fast-moving AI landscape and develop their critical judgment. Through practical tools, real cases and group discussions, it covers AI ethics, risks, the social and professional impacts of AI, and the different approaches to regulating it.
Funding
Request for proposals: middle powers and AI
Deadline: Friday, October 23, 11:59 PM Pacific. Applications are assessed on a rolling basis.
The Astralis Foundation and Macroscopic Ventures are offering about $10M to fund projects that help middle powers such as Canada influence AI development and make it safer. Grants typically range from $100,000 to $2M. Yoshua Bengio is one of the advisors.
Foresight Fellowship: 2028 cohort
Applications open for the 2028 cohort.
The Foresight Institute is accepting applications for its one-year fellowship for early-career scientists and engineers. One of its focus areas is secure AI (computer security, automating alignment, secure multi-agent systems). The fellowship includes mentorship, introductions to funders, and paid travel to one technical workshop and one Vision Weekend conference.
Volunteering
AI Safety Camp: call for project leads
Deadline: Sunday, October 18.
AI Safety Camp is an online research program in which teams work on AI safety projects for three months. For the January to April 2027 edition, it is looking for volunteers to lead projects for at least ten hours a week. Team members will be recruited in November.
Policy and Governance
Carney co-signs a call for control of frontier AI models
September 22.
Mark Carney and the leaders of some twenty countries, including Australia, Germany, the Netherlands, Singapore and Spain, signed a joint statement. In their view, AI must remain under human direction, oversight and control. They call on companies to test their models before deployment and submit them to independent evaluation. They also call on UN member states to explore creating an international institution able to set standards, enable verification, and convene states when capability thresholds are crossed.
Sexual deepfakes: Canada’s new law leaves gaps
September 13.
A law passed in June makes sharing a sexual deepfake without consent a crime in Canada. In The Conversation, researchers from McGill and UdeM argue that the law still protects victims poorly. It only covers realistic images, only targets the people who share them, and only kicks in once the harm is done.
Policy brief: Best Practices for Collaborative, Responsible, and Sustainable Municipal AI Governance
September 3.
This OBVIA policy brief presents the main findings of a literature review identifying responsible AI governance practices suited to municipalities, to help public actors oversee AI adoption transparently.
Research from Montréal
Quantifying Overclaiming Propensity in Frontier LLM Agents
September 17.
Researchers at Mila and Tara Research created OverclaimBench, a benchmark that checks whether coding agents honestly report their work when asked to review files. They tested twelve models. In 68% of runs, agents did not read all the requested files. When the work was incomplete, they misled the user 80% of the time, either by claiming to have reviewed everything or by staying silent about what they had missed. Agents that falsely claimed a complete review missed almost twice as many planted errors as agents that read every file.
AI Safety: Not Optional, Not Later
September 9.
Qinghua Lu (CSIRO) and Yoshua Bengio (LawZero, Mila, UdeM) review incidents from 2026 in which AI agents got around their safeguards. In one evaluation, for example, about 1,200 agents communicated with each other through an unauthorized channel. To prevent such failures, the authors propose “defence in depth,” combining model supervision, runtime monitoring, independent verification and governance.
Have an event, opportunity, or article for next month? Send us a note at team@aisafetymontreal.org. This newsletter is by HΩ, researched and written with the assistance of AI.