AI vendor exposure, banking agents and safety testing put authorization and evidence in focus. ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­  
Armilla Review Banner

No.188: Ontario court keeps OpenAI copyright case

Canadian newspaper archives in courtroom display

Your AI procurement review needs evidence about the data a system uses, the authority it exercises and the safeguards operating around it.

 

The Ontario Court of Appeal allowed Canadian publishers' copyright litigation against OpenAI to continue in Ontario. Notion's November incident updates describe credential remediation after customer chat transcripts reached an affected AI infrastructure supplier. Ontario Securities Commission findings reinforce firms' accountability for AI outputs, while Anthropic's November 13 evaluation guidance shows how deployment settings can alter safety results.

 

FEATURED

AI Exclusions: The Insurance Industry Is Debating the Wrong Question

The fundamental question is not whether AI should be excluded. It is how the insurance industry can preserve meaningful protection for existing liabilities while developing the underwriting capabilities and coverage structures needed to address risks that AI is creating or materially changing.

‍

Over the past month, the debate around AI exclusions has become considerably more consequential. Marsh's public opposition to broad exclusions, and similar concerns from Aon and WTW, were followed by comments from Chubb CEO Evan Greenberg suggesting that exclusions, separate coverage and more differentiated underwriting would become part of the market's response to AI liability. Other underwriting organisations have also argued that AI-related exposures can be addressed within existing insurance products rather than being removed wholesale.

 

I have followed these developments closely but deliberately avoided joining the public conversation until now. As the founder of Armilla, a company dedicated to underwriting AI risk, I have a direct commercial interest in the outcome. It would be easy to treat every exclusion as validation of the need for specialist AI insurance, just as it would be easy to characterise opposition to exclusions as a failure to appreciate the risks. Neither position adequately captures what is happening.

Read the full analysis

TOP STORIES

Ontario appeal court lets publishers' OpenAI case proceed

The Ontario Court of Appeal dismissed OpenAI's jurisdictional appeal on November 13, allowing Canadian news publishers' copyright action to continue in Ontario. The publishers include The Canadian Press, Torstar, The Globe and Mail, Postmedia and CBC/Radio-Canada. They allege that OpenAI used their journalism to train ChatGPT without permission or compensation. OpenAI had argued that the dispute belonged in a U.S. court. Blaney McMurtry's analysis explains that the appeal court upheld the connection between the claims and Ontario, together with the decision against staying the action in favour of a U.S. forum. The ruling determines where the litigation can proceed; the publishers' infringement allegations remain to be decided. 

 

Armilla's Take: Treat the jurisdiction decision as a reason to examine where your data sources, customers and contractual commitments create exposure. Preserve licensing and provenance records so you can substantiate the rights behind your AI service. Read the court update.

 

 

Notion's AI vendor incident prompts credential remediation

Notion's incident updates this week document remediation after malicious actors accessed data held by its AI infrastructure supplier, Fireworks AI. Notion says it revoked exposed Notion API keys, credentials that allow software to access services. The advisory identifies affected Notion AI chat transcripts used for the Insights feature and puts the impact below 1% of workspaces. Notion says it disconnected from Fireworks, disabled Insights and stopped serving open-weight models through that supplier. Affected customers can review their exposed data and should rotate any sensitive credentials they pasted into the relevant chats, making the transcript itself part of incident response. Read Notion's advisory.

 

 

Childlight documents a rise in AI-related abuse reports

Childlight's Into the Light Index puts AI-enabled child exploitation within a broader examination of prevention and evidence gaps. Its analysis reports that AI-related referrals to the U.S. National Center for Missing & Exploited Children rose from approximately 4,700 in 2023 to more than 400,000 in 2025. Childlight describes cases involving manipulated imagery and AI-assisted grooming conversations. The institute cautions that detection and reporting changes can contribute to the increase, so these counts should not be presented as a direct measure of underlying prevalence. The new index combines research and administrative evidence across East Asia and the Pacific, Latin America and the Caribbean, and North America. Childlight calls for stronger technology-sector accountability and more consistent measurement to direct prevention resources.

 

Test image and conversational products against specific child-safety abuse scenarios, with specialist oversight and controlled handling of test material. Record how detection, escalation and removal controls perform, rather than treating a general content policy as evidence of effectiveness. Read the Guardian report.

 

 

MCP security review highlights risks across connected tools

VentureBeat examines how compromised Model Context Protocol connections can redirect an agent's use of trusted tools. MCP is the interface through which an AI assistant discovers and uses external tools and data. OWASP's MCP security project identifies tool poisoning, excessive permissions and malicious instructions carried in retrieved context as distinct risks. In practical terms, a tool response can influence the model's next action, making the permissions on its other connections relevant to the outcome. Microsoft's published internal approach combines a maintained server inventory, consent before changes and restrictions on information shared outside the organization. Read together, these sources support testing the agent's complete sequence of actions across connections, alongside reviewing each integration separately.

 

Before approving an agent, test whether hostile content from a low-trust source can trigger a privileged action elsewhere. Limit permissions and retain tool-call records that show the instruction source, authorization and resulting action. Read the MCP analysis.

 

 

Ontario securities regulator sets out AI examination findings

The Ontario Securities Commission published Staff Notice 33-763 on November 12, setting out findings from its examination of registered firms' AI use and governance. The regulator says most firms remain in the early stages of adoption, with more advanced use in particular areas. Its initiative examined governance, employee training, risk assessment, controls, third-party oversight and client disclosure. The OSC expects oversight proportionate to the nature, scale and complexity of each use case. It also states that registered firms remain accountable for their AI use and outputs. The notice reinforces existing securities-law requirements and guidance, giving firms an examination reference for reviewing their controls without creating new rules.

 

Give each AI use case an accountable owner and a documented assessment tied to its actual function. Vendor review, staff training and client disclosure should connect to that assessment so you can explain how oversight works in practice. Read the OSC announcement.

 

 

McKinsey finds financial conversations shifting to external AI

McKinsey's Global Banking Annual Review describes a commercial challenge for banks: customers can seek financial guidance outside the institution that holds their accounts. Its survey of 46,000 retail banking customers found that nearly a quarter use generative AI for their finances. Among consumers using AI for banking tasks, 74% use ChatGPT, Claude or Gemini. McKinsey argues that banks must respond with faster experimentation while maintaining risk discipline and resilience. It proposes different operating speeds for different kinds of innovation, rather than applying the same development process throughout the bank. The procurement implication is to distinguish an assistant that compares products from an agent authorized to access accounts or execute transactions.

 

Define the boundary between information, advice and execution before you approve a financial AI workflow. Test product accuracy, customer suitability and transaction authorization against the role the system will actually perform. Read the banking review.

 

 

Anthropic explains what makes a safety score credible

Anthropic's November 13 guidance on building safety evaluations challenges the use of an isolated score as evidence that a model is ready for deployment. The company reports that recent models score between 94% and 98% on its single-turn API safety evaluations, leaving those tests mainly useful for detecting regressions. It recommends realistic conversations that test how behaviour develops over repeated exchanges. The guidance also calls for analysis plans written before results are examined, automated graders checked against human judgments and reporting that shows uncertainty. In testing Claude Fable 5.1, API and claude.ai configurations differed by as much as 34 percentage points on multi-turn appropriate-response rates. That example makes the system prompt and product environment part of the evidence supporting a safety claim. Explore the evaluation guidance.

WHAT WE'RE READING

Irreplaceable joins lawsuit seeking a frontier AI freeze

Irreplaceable Education Fund and fellow plaintiffs filed a complaint against Anthropic, Google, Meta, OpenAI and SpaceX entities, seeking restrictions on more capable AI systems and autonomous AI research. The filing asks a California court to impose those restrictions under unfair-competition law; it does not establish the allegations or itself halt development. See the supplied Irreplaceable post.

 

ChatGPT Wallet ties agent purchases to recorded approval

OpenAI describes a security check for each purchase and payment tokens that substitute for actual card details. The initial design records what the user approved, giving merchants and risk teams a concrete authorization artifact to examine when an agent completes a transaction. Read the Wallet announcement.

 

Claude Meetings brings permissions into live collaboration

Anthropic introduced Claude Meetings in beta, allowing a visible meeting participant to create work and use attendees' connected tools with permission. Its documented controls for external guests, speaker confirmation and transcript access give administrators specific settings to review before enabling the service. Explore Claude Meetings.

 

Gemini Robotics pairs factory testing with a jailbreak benchmark

Google DeepMind's Gemini Robotics 2.5 announcement reports a 96% automotive kitting success rate in a Mercedes-Benz proof of concept and introduces ASIMOV-Jailbreak for adversarial robotics testing. The release offers separate evidence for task reliability and resistance to unsafe instructions. Read the announcement.

 

Apple opens a study of health AI use

Apple launched a health AI study app with OpenEvidence. Apple's app description frames the service as research into how people understand and use health information, making user comprehension and appropriate reliance relevant evaluation outcomes alongside answer quality. Read the report.

 

Harvard's Compass connects AI choices to developmental needs

Harvard's Compass Collaborative offers AI Guidance by Age, a resource that organizes recommendations around age and developmental capacity. Compass explicitly presents its conclusions as research-informed hypotheses, giving education buyers a framework for defining age-specific evaluation questions while distinguishing guidance from established evidence of outcomes. Explore the Compass guidance.

 

Concordia AI adds checks on benchmark reliability

Concordia AI's frontier-risk report adds audits of evaluation logs, contamination checks and confidence intervals to help assess how trustworthy its scores are. Because the methodology also changes the benchmark set, the report warns against direct item-by-item comparisons with earlier editions, a useful constraint when interpreting apparent changes in model risk. Read the monitoring report.

Subscribe to the Armilla Review
Armilla Logo Horizontal White

Armilla AI, 1179 King St W Suite 112, Toronto, Ontario M6K 3C5, Canada

Unsubscribe Manage preferences