• NewCore emerges with $66M to give AI agents identities — As enterprises deploy thousands of autonomous agents, identity management becomes a critical bottleneck. NewCore's funding round signals that the "software worker" identity layer is emerging as the next infrastructure category for AI-native operations. 🔗 Graph: agentic-ai, ai-security, ai-governance, llm-gateway 📅 Published: 2026-06-15 📰 https://techcrunch.com/2026/06/15/ai-agents-are-becoming-employees-newcore-emerges-with-66m-to-give-them-identities/ 📌 Key takeaways: • NewCore builds an identity and access management layer specifically for AI agents — monitoring, authorizing, and revoking software workers that operate across enterprise systems without human oversight. • The company's CEO argues that identity is "one of the first enterprise systems strained by large-scale deployment of AI agents" — traditional IAM wasn't designed for non-human, programmatic actors running autonomous loops. • The $66M raise (led by prominent enterprise VCs) suggests the market views agent identity as a standalone category rather than a feature of existing IAM platforms like Okta or Azure AD. • For any organization deploying agentic workflows (including TritonAI's agent ecosystem), this directly impacts how API keys, OAuth tokens, and permission scopes must be managed as the number of software workers scales from dozens to thousands. • NewCore's approach could inform how the Developer API Program handles agent authentication on the LiteLLM gateway — separate from human user sessions and with revocation chains that don't depend on a single admin account.

• LiteLLM Vulnerability Chain: Critical exploit lets low-privilege users take over AI gateways — A three-CVE chain (combined CVSS 9.9) allows any default LiteLLM user to escalate to proxy_admin and execute arbitrary code on the gateway server. CISA has confirmed active exploitation of one vector. Teams running LiteLLM must patch to v1.83.14-stable immediately. 🔗 Graph: litellm-enterprise, ai-security, llm-gateway, enterprise-monitoring 📅 Published: 2026-06-16 📰 https://latesthackingnews.com/2026/06/16/litellm-vulnerability-chain-ai-gateway-patch/ 📌 Key takeaways: • The chain (CVE-2026-47101 → CVE-2026-47102 → CVE-2026-40217) starts with a route bypass that lets any user create a virtual key accessing admin endpoints, then self-promotes to proxy_admin via an unrestricted /user/update endpoint, then executes arbitrary code through LiteLLM's Custom Code Guardrail which runs admin-supplied Python via exec(). • A separate RCE in MCP REST test endpoints (CVE-2026-42271) is already in CISA's Known Exploited Vulnerabilities catalog — attackers weaponized it within 36 hours of disclosure. • Successful exploit hands attackers the master encryption key and all stored provider API keys, plus full access to intercept prompts and responses in flight — including any PII or proprietary model outputs passing through the gateway. • Partial patches are insufficient: versions below 1.83.14-stable may be vulnerable to one or more steps in the chain. The fix requires upgrading to the current stable release, rotating all provider API keys, and auditing the proxy_admin user list. • For UCSD's TritonAI deployment using LiteLLM Enterprise as the central gateway, this is a critical operational alert — the gateway sits between all campus AI users and 100+ model providers, making it the highest-value target on the infrastructure.

• The future of AI may be small, cheap and unprofitable — A recent study suggests small language models running on desktop computers may soon handle most tasks currently performed by large language models, challenging the "bigger is better" assumption that has driven the AI industry's massive compute investments. 🔗 Graph: model-agnosticism, vertical-ai, openai, google 📅 Published: 2026-06-18 📰 https://www.reuters.com/commentary/reuters-open-interest/future-ai-may-be-small-cheap-unprofitable-2026-06-18/ 📌 Key takeaways: • The study challenges the industry's central premise that scaling model size delivers proportional gains, suggesting diminishing returns for large-scale pretraining relative to smaller, task-specific models. • Desktop-grade inference could dramatically reduce the per-query cost of AI — from cents per API call to near-zero for local execution — which would fundamentally alter the economics of enterprise AI deployments. • For higher education, this could mean that the most cost-effective AI strategy is not renting cloud API access but running distilled or small models on existing campus infrastructure — directly relevant to UCSD's model-agnostic approach through LiteLLM. • The "unprofitable" angle in the headline points to a market tension: if small models commoditize inference, the profit pool shifts from API providers to the integration, governance, and orchestration layer — exactly where TritonAI is positioned. • Small models also change the security calculus: sensitive data never leaves the campus network, reducing the compliance surface for FERPA, HIPAA, and research data handling requirements.

• Azure Databricks at Data + AI Summit 2026: deep new integrations with M365 Copilot and Teams — Microsoft announced that Databricks Genie now works inside M365 Copilot Cowork, Teams threads, and Excel through Unity Catalog-governed data connections, plus a new SharePoint connector and OneLake catalog federation that reduce data movement between lakehouse and Microsoft ecosystem. 🔗 Graph: databricks, microsoft, llm-gateway, data-analytics, mcp 📅 Published: 2026-06-17 📰 https://techcommunity.microsoft.com/blog/azure-databricks/azure-databricks-at-databricks-data--ai-summit-2026-updates-and-new-announcement/4528388 📌 Key takeaways: • Genie for Microsoft Teams and M365 Copilot Cowork (Beta) allows users to tag Genie in a Teams thread and receive context-aware answers from the Azure Databricks lakehouse, with responses governed by Unity Catalog permissions — no SQL or dashboard required. • Copilot Studio agents can now reason over an entire Azure Databricks workspace through a single Model Context Protocol (MCP) connection — the same protocol that LiteLLM supports for agent tool routing. • The Azure Databricks Excel Add-in (Public Preview) brings governed lakehouse data into Excel without per-user ODBC setup, with write-back support and Unity Catalog metric views that keep business logic consistent across tools. • OneLake Catalog Federation (GA) enables bidirectional querying between Azure Databricks and Microsoft OneLake without data duplication — reducing the pipeline complexity that currently hinders cross-platform analytics. • For UCSD's data analytics strategy, these integrations mean that as the campus adopts M365 Copilot, the data governance framework (Unity Catalog) can extend into everyday productivity tools — faculty querying Tableau or Cognos datasets from within Teams without IT standing up new pipelines.

• Why University Classroom Technology is Now a Student Enrollment Strategy — Students are making enrollment decisions based on the quality and feel of campus learning environments, shifting classroom technology investment from a facilities cost to a core enrollment and retention lever for university IT leaders. 🔗 Graph: higher-ed-ai, ai-adoption, uc-san-diego, data-analytics 📅 Published: 2026-06-17 📰 https://edtechmagazine.com/higher/article/2026/06/university-classroom-technology-enrollment-strategy 📌 Key takeaways: • The EDUCAUSE 2025 Students and Technology Report found that students who perceive their institution as technologically cutting-edge report 85% satisfaction rates — versus institutions seen as lagging, where satisfaction drops dramatically. • CIOs now have a direct enrollment argument for classroom investment: institutions winning the enrollment conversation aren't just updating lecture capture but building complete tech-enabled environments (simulation labs, VR spaces, data science centers) that match where students expect to work. • The conversation has expanded beyond the classroom: students judge campus tech by game day experience, wayfinding, student unions, and athletics venues — expecting the same seamlessness they get from consumer services. • Three-phase framework identified: (1) students expect environments matching their ambitions, (2) classroom investment is now tied directly to enrollment and retention metrics, (3) the student experience extends well beyond the classroom into every campus touchpoint. • For UC San Diego and campuses like it, this reframes classroom IT budgets as retention infrastructure — provosts and boards now see the enrollment ROI case more readily than the traditional "modernize the AV system" pitch.

💡 Signal: Two themes dominated this week — the operational security of AI infrastructure (LiteLLM's critical CVEs, NewCore's agent IAM play) and the platform integration race (M365 Copilot + Databricks, small-model commoditization). Both point to the same inflection: the AI industry is shifting from building models to managing the secure, governed deployment of those models at scale — the exact territory TritonAI occupies in higher education.