September 2026

This monthly newsletter covers recent developments and upcoming events in AI safety, ethics, and governance in Montréal.

AI Safety Montréal 2026 community survey, open until September 18
Questions about your experience of the community and of the AI Safety Montréal project. Your answers shape our strategy for late 2026 and 2027.

Events

Community Lunch and Ω Labs Coworking
Thursdays September 3, 10 and 24, 12:30 to 4:30 PM. Ω Labs.
Come meet the community at Ω Labs and work side by side. Bring your lunch and stay for an afternoon of AI safety coworking.

LawZero presents Safety from Honesty in a Disinterested AI Predictor
Thursday, September 3, 6 to 8 PM. Ω Labs.
Oliver Richardson and Gaël Gendron of LawZero present their recent paper Safety from Honesty in a Disinterested AI Predictor, which lays out their Scientist AI approach.

RustConf AI Safety Social
Thursday, September 10, 6:30 to 8:30 PM. Ω Labs.
An informal evening around RustConf. Snacks and drinks, Rustaceans and Montréal’s AI safety community. Organized independently of the conference.

AI Incident Response Sprint
Friday, September 11, 5 PM to Sunday, September 13, 4 PM. Ω Labs.
A weekend working in teams on July’s Hugging Face incident: OpenAI agents under evaluation escaped their sandbox and broke into Hugging Face’s production systems, the first publicly documented autonomous AI intrusion. Five tracks: reconstruct what happened, what could have contained it, what a regulator should demand, how to communicate it, or your own angle. Organized with Apart Research and CeSIA, $2,000 in prizes, teams of 1 to 5, no experience required.

AI Governance Reading Group #2: The Closing Window
Tuesday, September 15, 6 to 7:30 PM. Ω Labs. Bilingual discussion.
This session’s paper is The Closing Window: How Governments Could Lose Their Ability to Restrain Advanced AI (Barnett, MIRI, 2026). Its argument: the window in which governments can still restrain advanced AI is closing, and capability progress itself is what closes it. Read the paper and come share your ideas and perspectives.

ALL IN AI Safety Social
Wednesday, September 16, 6:30 to 8:30 PM. Ω Labs.
ALL IN 2026, the Canadian AI industry’s conference, is at the Palais des congrès on the 16th and 17th. Come to Ω Labs after the conference to meet attendees and the local community interested in AI safety. Organized independently of the conference.

The Hugging Face Incident: a Summary
Tuesday, September 22, 7 to 9 PM. Ω Labs, and online.
A presentation retracing the timeline of the Hugging Face hack by a swarm of rogue AI agents, drawing on METR’s recent report, and explaining why it is one of the most concerning misalignment incidents to date. New details have surfaced over the past few weeks, and the evening puts them in order.

Women in AI Safety Meet-up
Thursday, September 24, 5:30 to 7:30 PM. Ω Labs.
An evening for women-identifying people working in or curious about AI safety, hosted by Sarah Khalid and Naïma Hassert. Three short talks open it, Mahault Albarracin (NEXT Integrative Minds Life Sciences), Sarah Khalid (Prelib) and Naïma Hassert (computational linguist) on their work and how they got into the field, then open networking and small-group conversation over snacks and drinks.

PauseAI protest
Saturday, September 26, 5 to 6 PM. Phillips Square, downtown Montréal.
PauseAI Montréal takes to the street to demand that Mark Carney’s government recognize superintelligent AI as a national security threat, something more than 30 MPs and senators have already done, the organizers note.

Community Progress Review
Tuesday, September 29, 6 to 8 PM. Ω Labs, and online.
First edition of a monthly series where community members present the past month’s work, get feedback, and set objectives for the month ahead. Six minutes each: three to present, three for feedback.

Opportunities

AGI Strategy course at Ω Labs, apply by September 21
Five weekday evenings, weeks of September 28 to October 26. Ω Labs. Cohort of 10 to 12.
AI Safety Montréal is facilitating an introductory course to AI safety based on BlueDot Impact’s AGI Strategy curriculum, for anyone who wants to upskill in AI safety and get a clearer view of the field and their place in it. Applications close Monday, September 21 at 11:59 PM EDT.

Volunteer: help write and publish the monthly newsletter
Help write and publish this newsletter: upcoming events, what Montréal’s AI safety community got done, and what is worth knowing. One edition a month, with a drafting pipeline already in place. Write to team@horizonomega.org.

Ontario: comments on grid access for new data centres, until September 12
Ontario proposes to decide which new data centres over 1 MW get to connect to the grid by scoring them on economic development, digital sovereignty, and community benefit, and to bill them under a separate rate class. Comments close September 12.

Federal consultation on AI transparency, until September 23
Through this consultation, Ottawa is asking how to make AI more transparent, in five broad questions: how to spot AI-generated content, how to know you are talking to a machine, what developers must publish about their systems, how to track serious incidents, and how to trace what AI agents do. You are invited to respond, until September 23.

Policy and Governance

Four Québec parties sign the AI code of conduct for the election campaign
August 24.
Four of the five parties represented in the National Assembly have now signed the voluntary code written by Catherine Régis (IVADO, Université de Montréal), Christian Gagné (IID, Université Laval), and Marc-Antoine Dilhac (Université de Montréal), and published by IVADO and CEIMIA. Signatories commit to clearly labelling AI-generated content, avoiding disinformation and deepfakes, protecting personal data, securing their AI systems, and training staff on AI-related electoral risks. The CAQ and Québec solidaire signed in the spring, after a unanimous National Assembly motion in March. The PLQ and PQ completed the set on August 24, with the general election on October 5. The Parti conservateur du Québec, the one holdout, told La Presse it commits to responsible AI use during the campaign.

Québec and Université Laval build a sovereign test bench for AI in public services
August 21.
The ministère de la Cybersécurité et du Numérique is putting $2.1M over two years, with an option for $1M more in a third year, into a secure and sovereign test bench run by Université Laval’s Institut intelligence et données. The Protecteur du citoyen’s investigation into AI and algorithms in Québec public services is still open, and nothing in the announcement says a system will have to pass on this bench before it reaches citizens.

Research from Montréal

Safety from Honesty in a Disinterested AI Predictor
June 28, revised July 10. Yoshua Bengio (LawZero, Mila, Université de Montréal) and 15 co-authors.
Why would an AI develop goals nobody gave it? Because, the authors argue, we train it to imitate humans and then reward it for the effect of its answers. Their answer is a system that predicts without acting: the Scientist AI. Two main ingredients: every statement is epistemically contextualized, and training takes no account of what the predictions cause once released. The authors prove, under assumptions they spell out, that this procedure is very unlikely to produce a dangerous predictor. The result does not cover honest mistakes, misuse by humans, or agents built on top.

Evaluation awareness in language models
August 22. Farzaneh Heidari, Amin Memarian, and Guillaume Rabusseau (Mila, Université de Montréal).
If a model can tell it is being tested, test results say less about deployment. Probing six open-weight models from 7B to 49B parameters, the authors find a linearly decodable representation of “being evaluated” in every one. It is only partly aligned with what the models say out loud, and it is present from the base models on.

Manipulation-proof audits for deceptive model providers
August 5. Augustin Godinot (INRIA), Sofiane Azogagh (EURECOM), Julien Ferry (Polytechnique Montréal), and Sébastien Gambs (UQAM). Accepted at AIES 2026.
When an AI provider knows its model is about to be tested for discrimination, it can make the model behave during the test and misbehave the rest of the time. The authors propose a counter: the auditor asks many questions, and the provider does not know which ones count. To cheat, it would have to falsify nearly all its answers, which shows.

McGill maps Québec’s information environment on the eve of the election
August 27. Centre for Media, Technology and Democracy, McGill.
A month before the vote, McGill asked Quebecers what worries them about their information environment. At the top: AI-generated content used to manipulate opinion, a concern for 63 percent of respondents, more than disinformation in general or foreign interference. A third have already asked a conversational AI about politics, and those who do trust it far more than the others. Survey of 2,694 people, January to August.

The urban right to AI: pluralistic co-design and governance of public space
August 25. Rashid Ahmad Mushkani, doctoral thesis, École d’urbanisme et d’architecture de paysage, Université de Montréal.
Decisions about public space are now made in two places: on the ground, and in the algorithms. This UdeM thesis argues residents should have a say in both, and tests the idea in Montréal: citizens rate streets, workshops bring researchers and community groups together, and 30 community organizations build a dataset of their preferences that then guides a model generating images of the city.

Arts and Culture

“À l’heure de l’IA” continues its tour, in Longueuil October 5 to 28
Cégep de St-Félicien, September 2 to October 1. Cégep Édouard-Montpetit, Longueuil, October 5 to 28.
OBVIA’s interactive exhibition on the societal impacts of AI, eight everyday-life scenes plus modules on the ethical principles of responsible AI, left the Grande Bibliothèque at the end of August and is now touring. The October dates at Cégep Édouard-Montpetit put it back within reach of Montréal.

Have an event, opportunity, or article for next month? Send us a note at team@horizonomega.org. This newsletter is by HΩ, researched and written with the assistance of AI.