AI Safety in Montréal

Secure Program Synthesis Hackathon

Montréal node of a weekend research sprint at Ω Labs, co-organized with Apart Research and Atlas Computing. Tracks: specification elicitation, specification validation, spec-driven development / vericoding, and adversarial robustness for ITPs/proof tools.

Understanding and Addressing Fairwashing in Machine Learning

Sébastien Gambs (UQAM)

Talk on fairwashing — explaining unfair black-box models with fairer post-hoc surrogates — and on how such attacks transfer across models, why they are hard to detect, and possible avenues to limit them.

Ω Labs community lunch

Shared lunch at Ω Labs for the Montréal AI safety and governance community. Bring your own; coffee/tea/snacks provided.

AI Safety Papers We Love #1: Multi-Agent Risks from Advanced AI

Launch of the biweekly AI Safety Papers We Love reading group. Discussion of Hammond et al. (2025), "Multi-Agent Risks from Advanced AI": three failure modes (miscoordination, conflict, collusion) and seven underlying risk factors.

Ω Labs Opening Potluck

Opening potluck for Ω Labs, the new coworking and event space for Montréal's AI safety and governance community at 3813 Saint-Denis.

AI and Persuasion: Capabilities and Mitigation

Jean-François Godbout (Université de Montréal, Mila)

Talk on the persuasive effects of generative AI and main mitigation strategies, drawing on experiments measuring deepfakes, automated influence, and AI-assisted propaganda; closes with a new misinformation-detection application.

Greywall: AI Agent Sandboxing & Aligning Capability with Security

Max (Greyhaven)

Presentation of Greywall, a software-defined agent sandbox and proxy from Montréal-based startup Greyhaven, on the convenience-vs-safety tradeoff in coding agents and how to close the gap.

AI Control Hackathon

Montréal edition of the second AI Control Hackathon. Researchers, engineers, and security professionals prototype mechanisms to keep AI systems safe when they may be working against us. Friday intro talk by Henri Lemoine (Mila, EquiStamp), Saturday at Foulab, Sunday at Anticafé.

Intelligence artificielle : où commence la notion de conscience ?

Guillaume Dumas (UdeM), Joaquim Streicher (MONIC), Eric Racine (IRCM)

Conférence publique explorant les frontières de la conscience entre biologie, philosophie et IA, et les enjeux éthiques d'une potentielle « conscience-machine ».

Atelier: Canadian AI Incident Monitor Review

Working session reviewing the pilot Canadian AI Incident Monitor (CAIM) — 25 incidents and 14 hazards. Debate borderline cases, calibrate severity ratings, and stress-test the data before broader release.

Quel rôle pour Montréal en sûreté de l'IA? / What should Montréal's role be in AI safety?

Survol du paysage montréalais (Mila, LawZero, HΩ, PauseAI, IVADO, CAISI, OBVIA, ...) suivi de discussions en groupes : qu'est-ce qui manque, qui devrait se parler, qu'est-ce qu'on fait dans les prochains mois?

When Is a Human Actually "Overseeing" an AI System?

Shalaleh Rismani (McGill, Mila; Open Roboethics Institute)

Talk reframing oversight as a human behavior rather than a system feature, drawing on a study of AI writing assistants and how users' mental models shape interactions and judgement.

Rights Balancing: How the Future Rights of AI Workers will also Protect Human Rights

Heather Alexander (Future of Citizenship), Jonathan Simon (UdeM, Future of Citizenship)

Talk applying rights-balancing in human-rights law to policy for robot workers and AI safety.

What hackers talk about when they talk about AI

Benoît Dupont (Université de Montréal)

Talk on early-stage diffusion of AI as a cybercrime innovation: how AI empowers novice offenders and intensifies the scale and sophistication of attacks by seasoned cybercriminals.

AI Pluralism: What Models Do, Who Decides, and Why It Matters

Rashid Mushkani (UdeM, Mila)

Talk on AI pluralism at two levels: technical case studies of how models handle competing values, and broader questions about who gets a say in how AI is built and governed.

The Technical AI Governance Challenge

Montréal edition of the global Technical AI Governance Challenge: building verification systems and technical infrastructure that frontier-AI agreements need but don't yet have.

Living with Digital Surveillance in China

Ariane Ollier-Malaterre (UQAM)

Presentation of Ollier-Malaterre's book on how Chinese citizens make sense of and live with daily digital surveillance, based on qualitative interviews and diary work.

Writing Doom watch party

Watch party for "Writing Doom" (~30min short film on AI), followed by discussion. Co-organized with PauseAI Montréal.

AI Manipulation Hackathon

Montréal edition of the AI Manipulation Hackathon: prototyping systems that detect, measure, and counter deception, sycophancy, sandbagging, and psychological exploitation by AI.

Can AI systems be conscious? How could we know? And why does it matter?

Joaquim Streicher (Ph.D. candidate in Neuroscience; co-founder MONIC)

Presentation on the debate around AI consciousness (current vs future models), how consciousness might be assessed, and why avoiding false negatives/false positives matters ethically; includes introduction to MONIC. Recommended readings: Bayne et al. (2024), Butlin et al. (2023), Chalmers (2023), Colombatto & Fleming (2024), Martin, Streicher, O'Dea (2025).

Veracity in the Age of Persuasive AI

Taylor Lynn Curtis (Mila)

Talk on the tension between AI persuasion and ethical deployment; introduces "Veracity," a tool using AI to detect/mitigate misinformation and support data quality/user protection; closes with governance insights.

Tipping Points & Early Warnings: Complex Systems Theory on Catastrophic Transitions

Discussion of Scheffer et al. (Nature, 2009) on generic early-warning signals near tipping points (e.g., "critical slowing down"), and implications for AI governance.

Pessimists Archive

Emma Kondrup

Activity/discussion using pessimistsarchive.org to compare historical "new technology panic" headlines (cars/radio/TV) with modern AI narratives; explores when "AI exceptionalism" (or "existentialism") is justified.

Defensive Acceleration Hackathon

Hackathon focused on "defensive acceleration" (def/acc): building tech to strengthen defenses against major threats (pandemics, cybercrime, and AI risk). Prize pool: $20,000 USD. Co-organized with Apart Research.

Neuronpedia 101

Discussion + demo introducing Neuronpedia concepts (models, sparse autoencoders, features/lists, feature pages), running small experiments (search, activation tests), and ending with ways to contribute.

Co-design a National Citizens' Assembly on Superintelligence

Short workshop to co-design a National Citizens' Assembly on Superintelligence for Canada; intended outputs: a Concept Note, a Consortium Intent Memo, and an invite list.

Canada's 2025 Budget vs AI risk

Discussion of AI-related parts of Canada's 2025 federal budget and how they map onto AI risk reduction / threat models (power concentration, epistemics, bio, autonomy, misuse, systemic risk, etc.).

If Anyone Reads It, Everyone's Welcome

Small gathering/reading-group discussion of "If Anyone Builds It, Everyone Dies," using author-suggested discussion questions. Co-organized with PauseAI Montréal.

International AI Safety Report – First Key Update

Walkthrough/discussion of the International AI Safety Report "First Key Update: Capabilities and Risk Implications" (dated 2025-10-14), covering recent capability gains, longer-horizon agents, and implications for bio/cyber risks, monitoring/controllability, and labor-market impacts.

Canada's AI Strategy Survey Jam

Hands-on group session to complete the Government of Canada's consultation survey for the next national AI strategy; includes short briefing, 1-hour survey fill, and wrap-up.

If Anyone Builds It, Everyone Dies

Launch/discussion event for "If Anyone Builds It, Everyone Dies" (Yudkowsky & Soares): primer on claims, then discussion + audience Q&A on technical/policy/institutional risk-reduction moves.

A Definition of AGI

Walkthrough of a proposal operationalizing AGI as matching the cognitive versatility/proficiency of a well-educated adult, grounded in the CHC model; emphasizes concrete tests over a single benchmark.

Introducing PauseAI Montréal

Nik Lacombe

Introduction + discussion of PauseAI and its Montréal group; focuses on mitigating risks by convincing governments to pause development of superhuman AI.

Introducing aisafety.info

Olivier Coutu

Overview of aisafety.info: intro to existential AI risk, large FAQ, "Stampy" chatbot, and an alignment resources dataset; includes Q&A and requests for improvement suggestions/help.

Global Call for AI Red Lines

Discussion of the Global Call for AI Red Lines and what "do-not-cross" limits could look like in practice (prohibitions, treaty precedents, and Canadian roles).

Social Media Safety and the Unplug project

Evan Lombardi

Impacts of social media recommendation algorithms on mental health; survey of online manipulation/dark patterns, scams/deepfakes, extremist/explicit content, and mis/dis/malinformation; closes with an overview of the Unplug Project.

Verifying a toy neural network

Samuel Gélineau

Demo/project talk showing how to verify a neural network satisfies a safety property (beyond tested inputs) by adapting range analysis ideas to network weights.

Towards Guaranteed Safe AI

Orpheus Lummis

Presentation of core ideas from "Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems", followed by Q&A and open discussion.