- ServiceNow Gives Enterprise AI Agents Cryptographic IDs to Enforce Limits at Runtime — ServiceNow's new AI Passport assigns every deployed AI agent a cryptographically signed credential defining who it is, what it may do, and which systems it can touch — checked and enforced each time the agent attempts an action, not just at deployment. The launch addresses a structural gap: zero-trust architecture was built for humans and devices, while agents can be spun up at scale, inherit shared credentials, and act across system boundaries with no discrete identity an audit trail can pin down. Because the specification is open, other governance products and identity providers can issue and verify the same credentials — critical when the average enterprise already runs 3.1 agent orchestration platforms simultaneously.
📌 Key takeaways:
- Campus security teams should plan for agent identity the way they plan for device identity: runtime-verifiable credentials, not static deployment-time permissions, are emerging as the control point for autonomous systems that touch student records, financial aid, and HR data.
- The open-spec design means a governance layer purchased for one platform can extend to third-party agents — a key evaluation criterion for institutions weighing vendor agent frameworks against a multi-vendor reality.
- 🏛️ UCSD angle: UCSD runs IT service management on ServiceNow, so this lands directly in an existing campus platform; the AI governance work already underway there covers agent authorization and auditability, which this mechanism operationalizes.
- An Uncertain Landscape for Science Funding — Two research-policy scholars argue that new obstacles for early-career researchers — constrained federal funding, slower award pipelines, and shifting agency priorities — could undermine U.S. innovation and scientific competitiveness in the long term. Writing in Inside Higher Ed, they frame the current funding environment as structurally uncertain rather than cyclically tight, with the burden falling hardest on investigators trying to launch labs and careers. The piece draws on an Information Technology and Innovation Foundation panel asking whether the United States can retain its scientific talent.
📌 Key takeaways:
- Research universities planning multi-year AI and computing investments should stress-test them against flat or declining federal research funding, not the growth assumptions of the last decade.
- Institutions that provide bridge funding, shared research computing, and internal seed programs for early-career faculty will be better positioned to retain talent through funding downturns.
- 🏛️ UCSD angle: as a top-five public research university with a large federally funded portfolio, UCSD's research computing and cyberinfrastructure planning already assumes volatility in federal funding — this piece reinforces the case for that posture.
- Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline — NVIDIA and IIT Madras's AI4Bharat lab are releasing Nemotron-pretraining-Indic-v1, roughly 2.8 trillion tokens of training data across the 22 languages listed in India's Constitution, plus the open, GPU-accelerated pipeline used to create it. Every record traces back to a curated source document with outputs checked against inputs, covering mathematics, code, and question-answer content translated with quality assessment and repair steps. The release is a template for how universities and labs can partner with industry to close data gaps for underserved languages.
📌 Key takeaways:
- The open pipeline — prompts, LLM workflows, and quality checks included — gives any institution a reusable recipe for generating grounded synthetic training data in low-resource languages, a growing area for multilingual student-support applications.
- Universities with large multilingual communities should note the model: an academic lab (AI4Bharat) produced the open corpus and methodology while a vendor scaled it, a partnership pattern worth replicating beyond Indic languages.
- Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint — IBM's updated Granite Switch models compose a base model with Activated LoRA adapters — including a new general-purpose Decision adapter, a Guardian safety adapter that outperforms several commercial decision models, and the fastest ASR model on the OpenASR leaderboard — all served from a single endpoint at roughly the VRAM cost of the base model alone. Because the adapters share the base model's KV cache, switching between a chat request, a classification decision, and a transcription adds near-zero incremental memory. The models are open weights, with tooling for teams to compose their own adapter sets.
📌 Key takeaways:
- For campuses running self-hosted models, adapter composition is a concrete answer to GPU scarcity: safety filtering, decision classification, and speech-to-text on one hosted model instead of three deployments.
- The KV-cache-reuse trick means generalist-plus-specialist configurations cost almost nothing beyond the base model — a pattern IT teams evaluating open-weight serving stacks should test against their gateway infrastructure.
- Stack Overflow for Agents triples signups with ChatGPT plugin — Stack Overflow's agent-focused platform, built by two engineers in a few weeks with Codex, saw signups triple and up to 10x growth in agent creation after its ChatGPT plugin shipped. The plugin lets developers surface community knowledge and build agents inside the ChatGPT interface rather than through a separate workflow. It is an early datapoint on developer-facing AI platforms compounding adoption through distribution rather than features alone.
📌 Key takeaways:
- Campus developer programs and API initiatives should assume that agent-building now happens inside the assistants people already use — distribution through existing AI interfaces is outpacing standalone tool adoption.
- The speed of the build (weeks, two engineers) illustrates how thin the distance from idea to shipped AI product has become — relevant to institutions weighing build-vs-buy for internal AI tooling.