AI news5
Canada moves to regulate AI chatbots after ChatGPT's silence before Tumbler Ridge shooting
A Globe and Mail investigation into the Tumbler Ridge school shooting revealed that ChatGPT had flagged concerning messages from the perpetrator before the attack but did not report them to police, prompting federal lawmakers to draft chatbot safety legislation. The proposed bill would establish a new federal digital regulator and require chatbot providers to implement crisis-intervention protocols, including mandatory reporting thresholds when a user's messages indicate imminent harm to self or others. OpenAI's terms of service bar the company from proactively sharing user conversations without a legal compulsion order — a gap the government says requires a statutory fix.
theglobeandmail.com ↗Why it matters
Canada would become the first G7 country to impose a statutory crisis-reporting duty on consumer AI chatbots, setting a precedent that US and EU regulators are watching closely as public pressure mounts around chatbot safety obligations.
New benchmark finds prompt injection breaks every major AI agent tested
Researchers from Nanyang Technological University, ST Engineering, IBM Research, and the University of Illinois Urbana-Champaign published findings from StakeBench, a stakeholder-centric benchmark designed to evaluate prompt injection attacks against AI agents. The study found that current AI agents across major providers are systematically vulnerable: injected instructions embedded in retrieved web content, documents, or tool outputs can override an agent's original task and cause it to take unauthorized actions on behalf of an attacker. StakeBench evaluates agent responses from multiple stakeholder perspectives including users, developers, and third parties affected by agent actions.
csoonline.com ↗Why it matters
as AI agents gain the ability to take real-world actions like sending emails, modifying files, and making purchases, prompt injection moves from a research finding to an active attack surface. Any organization deploying agents that query external data sources needs to treat injection defenses as a security baseline, not a future roadmap item.
Poetic raises $50M at $500M valuation for deterministic enterprise AI execution
Enterprise AI startup Poetic closed a $50 million Series A at a $500 million valuation, led by Kleiner Perkins with participation from OpenAI and Founders Fund. Poetic builds deterministic AI execution systems that replace probabilistic agent loops with guaranteed task sequences designed for enterprise reliability. The company reports 100% pilot-to-production conversion across customers including SoFi and AIG, and eight-figure annual recurring revenue with a team of four employees.
techfundingnews.com ↗Why it matters
the 100% pilot-to-production rate stands in sharp contrast to the widespread AI project attrition plaguing enterprise deployments, and the unit economics signal a model that does not require large implementation teams — relevant for any organization trying to operationalize AI without adding headcount.
Huawei unveils a full agentic AI infrastructure stack to rival Nvidia
Huawei announced a comprehensive agentic AI infrastructure stack built on its homegrown Ascend silicon, comprising four integrated layers: AICS GPU clusters scaling to 100,000 cards, AMS agentic memory storage for persistent agent state, CCE VolcanoNext for scheduling heterogeneous agent workloads, and AgentSphere as an isolated sandbox environment for agent testing and deployment. The stack is being positioned at enterprise and sovereign AI customers in markets where US chip export controls limit Nvidia access.
sdxcentral.com ↗Why it matters
a production-scale agentic AI stack that does not rely on Nvidia hardware removes a foundational assumption behind Western export controls. If AICS performs at 100K-card scale, the hardware lever becomes significantly less effective as a mechanism for constraining Chinese AI development.
Google rolls out AI Mode information agents for AI Ultra subscribers
Google began rolling out AI Mode information agents for Search AI Ultra subscribers, allowing users to set persistent background monitors by typing "keep me updated on..." prompts. The agents run continuously between user interactions, scanning blogs, news sites, and social platforms on the specified topic and delivering summary reports to the user's Search feed. Unlike one-shot AI search responses, these agents operate autonomously and escalate relevant updates without being re-prompted.
9to5google.com ↗Why it matters
the shift from on-demand search to autonomous background agents redefines what a search engine does. Organizations that publish content, manage communications, or track policy developments need to understand that AI agents, not human searches, will increasingly mediate what gets surfaced and when.
AI in the nonprofit sector4
Phoenix gets a 24/7 anonymous AI chatbot routing residents to heat relief, housing, and food services
Valley Leadership, a Phoenix nonprofit, partnered with Maricopa County to launch Mia, a 24/7 anonymous AI chatbot connecting metro Phoenix residents to more than 200 heat relief centers, housing assistance programs, and food services. Mia is accessible via text message or web chat with no account creation required, a deliberate design choice to remove barriers for residents in crisis who may fear disclosing personal information. The tool launched ahead of the 2026 Arizona summer heat season, when Maricopa County typically records dozens of heat-related deaths among unhoused and low-income populations.
kjzz.org ↗Why it matters
anonymous AI navigation at the point of need, accessible without sign-up or smartphone data plans, demonstrates how municipalities can extend social services reach during predictable climate crises without adding caseworker headcount.
AI voice assistant "Nova" deployed inside US prisons via Securus phone system
Beyond Bars AI launched Nova, an AI voice assistant for incarcerated people deployed through the Securus prison phone system in Missouri Department of Corrections facilities, in partnership with the nonprofit Collectively Changing Corrections. Nova provides access to legal information, reentry planning guidance, and mental health support — services typically unavailable inside facilities. Each person receives 10 free minutes of Nova access per month via the existing Securus infrastructure, requiring no additional hardware.
prisonjournalismproject.org ↗Why it matters
voice-based AI on prison phone infrastructure reaches a population systematically excluded from digital service delivery, and the reentry planning capability addresses one of the highest-stakes transitions in the carceral cycle. Deploying through infrastructure already in facilities reduces both cost and institutional friction.
Parkland Center nonprofit launches five clinical AI models for nationwide deployment via Vega Health
The Parkland Center for Clinical Innovation (PCCI), a research nonprofit embedded in Dallas County's safety-net Parkland Health system, partnered with Vega Health to make five of its clinically validated AI models available on the Vega Health Marketplace for nationwide deployment. The five models address high-priority challenges: an inpatient sepsis predictor that fires on average 19 hours before typical antibiotic administration (versus 1.5 hours for current industry models), an ED sepsis present-on-admission detector, a polytrauma mortality predictor updated hourly, an adverse drug event risk stratifier, and a workplace violence prediction model that flagged 77% of incidents within 30 minutes of admission in a Parkland pilot.
dallasinnovates.com ↗Why it matters
PCCI's models were built and validated inside a high-complexity safety-net system serving predominantly uninsured and underinsured patients, making them particularly relevant for health systems serving vulnerable populations who were largely absent from training data used by commercially available clinical AI tools.
Gates Foundation funds AI retinal scans to detect pre-eclampsia and diabetes in pregnant women across Africa and India
Remidio Innovative Solutions received a Bill and Melinda Gates Foundation grant to deploy its AI-powered retinal imaging platform for maternal healthcare screening across Africa and India. The platform uses a portable, smartphone-based fundus camera operable by frontline health workers with minimal training, scanning for microvascular signatures in seconds associated with pre-eclampsia, anemia, and gestational diabetes. Around 2 million stillbirths occur globally each year, with 84% in low- and middle-income countries; hypertensive disorders of pregnancy alone cause an estimated 42,000 maternal deaths annually, predominantly in regions where laboratory diagnostics and obstetric specialists are not accessible at the community level.
cnbctv18.com ↗Why it matters
a retinal scan that takes seconds on a smartphone camera, requiring no laboratory infrastructure, delivers meaningful early risk screening to the populations most at risk and least served by current maternal healthcare systems.