- OpenAI Introduces Dots, AI Agents that Continue Working Between User Conversations — OpenAI has announced Dots, GPT-6 Astra-powered agents that run on their own cloud computers, use connected applications, and carry on with assignments between conversations while incorporating user feedback. OpenAI is also previewing specialist Dots for organizations — with separate identities and credentials for defined responsibilities — beginning with focused pilots, and is working with Microsoft to integrate with Agent 365's governance and security controls. The shift from chat sessions to persistent delegation changes the supervision math: an instruction that was appropriate yesterday may no longer fit today, and OpenAI's own testing found scope violations increased as intervening tasks accumulated.
📌 Key takeaways:
- Persistent agents separate two lifecycles IT leaders must govern independently: disconnecting an app stops new access but does not remove information already incorporated into the agent's context, so offboarding an agent requires deciding what it already knows.
- Campus security and governance teams should treat always-on agents as a new identity category — with credential lifecycles, scope reviews, and explicit stop conditions — before faculty and staff begin connecting institutional apps to them.
- Availability lands in education workspaces as an administrator-enabled beta, which makes the workspace admin the de facto policy gate: institutions should decide their enablement stance now rather than by exception later.
🏛️ UCSD angle: UCSD's centralized LLM gateway and AI governance work give the campus a natural control point for persistent agents — agent traffic and model policy can be enforced at the gateway layer before agents reach institutional data.
- Survey Reveals Data Governance Enforcement Gap for Agentic AI Tools — A new Delinea report finds that while 99.7% of surveyed IT and security leaders say their organization has a formal policy governing what data AI tools and agents may access, only about 51% said AI access is checked against those policies in real time — and only 19% could detect in real time when an agent accessed data outside its intended scope. The survey of 2,254 IT/security leaders and 2,250 employees also found 87% of organizations experienced or suspected an AI-related incident in the past year, and that employees routinely bypass AI approval processes. The report's title — "The AI Enforcement Gap" — captures the pattern: governance on paper is outpacing the technical controls needed to enforce it.
📌 Key takeaways:
- Having an AI data-access policy is now near-universal; enforcing it is not — institutions should audit the distance between their written AI policies and what their identity, logging, and access tooling can actually detect in real time.
- AI agents inherit employee permissions and retain access after tasks end, which makes periodic agent access reviews as necessary as privileged-account reviews.
- Employee bypass of AI approval processes means governance programs that rely on voluntary intake will undercount shadow AI use; discovery and network-level visibility matter more than policy attestations.
🏛️ UCSD angle: UCSD's gateway-based AI architecture already centralizes logging and model policy enforcement — the exact real-time enforcement layer the survey finds missing at most organizations.
- GPT-6 and Intelligent UI for everyone — OpenAI is rolling out GPT-6 globally in ChatGPT with Intelligent UI, which combines text, visuals, and interactive elements to fit the question — side-by-side comparisons, diagrams, and in-conversation tools such as calculators and bill splitters — choosing the format automatically and falling back to plain text when that works best. GPT-6 can also begin answering while it continues thinking or using tools, then append findings without another prompt. Paid and business tiers get GPT-6 Sol, free tiers get GPT-6 Luna, and the rollout to free users follows a day after paid plans.
📌 Key takeaways:
- The free-tier rollout puts a frontier multimodal interface in every student's pocket within days, which resets the baseline for what campus AI literacy training and acceptable-use guidance need to cover this term.
- Streaming answers that continue after the first response complicates simple "verify the output" guidance for students and staff — instruction should address partially-delivered, evolving answers.
- For institutions negotiating enterprise AI contracts, model tiers now differ by plan (Sol vs. Luna vs. Astra), so "ChatGPT access" alone is no longer a well-defined procurement line item.
🏛️ UCSD angle: UCSD's AI platform is deliberately model-agnostic, routing many providers through one gateway — new default models like GPT-6 get evaluated and provisioned centrally rather than locking the campus into a single vendor's rollout cadence.
- What IT Leaders Need To Know Before Deploying a Campus Mental Health Platform — As 24/7 online mental health platforms gain traction in higher education, EdTech Magazine walks through what they mean for IT leaders: these tools handle protected health information, and the technology review process that institutions apply to ordinary SaaS isn't sufficient on its own. The piece argues that evaluating these platforms requires pulling in counselors and HIPAA compliance officers alongside standard security and privacy reviews, working closely with vendors to set expectations, and understanding risk profiles before student data flows.
📌 Key takeaways:
- Campus IT teams procuring always-on mental health platforms are effectively taking on HIPAA-scope PHI management, which should trigger an expanded review that includes clinical and compliance staff, not just the standard security questionnaire.
- The AI components of these platforms (triage chat, risk flagging) make vendor questions harder: institutions should ask what the model does with a student in crisis, who is notified, and what the escalation path to a human is.
- Contract terms should specify data segmentation, breach notification, and what happens to student records if the platform relationship ends.
- Survey: Lack of Specialized Infrastructure Is a Top AI Adoption Challenge — A Digital Realty survey of 2,131 IT decision-makers across 19 countries finds 40% citing lack of specialized AI infrastructure as a main barrier to adopting a formal AI strategy — up from 9% in the firm's 2024 research. Half of respondents host AI models in private clouds, 92% tie data-location strategy to their AI plans, and 86% are pursuing or exploring sovereign AI. Organizations are also shifting from planning to execution: the share actively executing AI for operational efficiency rose from 27% in 2024 to 37%, with 63% expecting ROI within six months to two years.
📌 Key takeaways:
- Infrastructure has become the leading AI adoption constraint — ahead of regulatory and data-quality concerns — which means campus AI strategies without a facilities, power, and network component are incomplete on paper.
- The 92% tying data-location strategy to AI plans reflects where higher ed already lives: research data residency, FERPA, and state policy all force explicit placement decisions that commercial surveys are only now catching up to.
- For colleges without capital for dedicated AI infrastructure, the private-cloud-plus-provider-API hybrid most respondents describe is a realistic template — on-prem where data sensitivity demands it, hosted models everywhere else.
🏛️ UCSD angle: UCSD's AI platform already runs the hybrid pattern the survey describes — on-prem research compute alongside cloud model access brokered through a central gateway — keeping data-placement decisions explicit rather than accidental.
- Now in Nature: Retrofitting language models to operate over bytes — The technique behind Bolmo, Ai2's fully open byte-level language models, is now published in Nature, with new checkpoints showing the approach generalizes beyond the original Olmo model family. Byte-level operation removes the tokenizer layer entirely — historically a barrier for multilingual and low-resource language applications — and the open release includes checkpoints, code, and research details.
📌 Key takeaways:
- Tokenizer-free architectures matter for universities with multilingual student populations: byte-level models avoid the vocabulary biases that degrade performance for non-English languages, a recurring equity issue in campus AI deployments.
- A fully open model family with a Nature-validated technique gives research universities a self-hostable option where data sensitivity or cost rules out commercial APIs.
- The release pattern — open checkpoints plus formal peer review — is a template for how university research offices should expect open model research to be published and evaluated.
🏛️ UCSD angle: UCSD's gateway can route open, self-hostable models alongside commercial ones, so tokenizer-free open models like Bolmo are a practical addition to the campus model menu rather than a curiosity.