- Survey: Half of Companies Confirm Unauthorized AI Agent Actions — Nearly half of surveyed organizations (49.8%) confirmed or suspected that an AI agent took an action they did not intend, expect or authorize in the past year, according to Salt Security's State of Agentic AI and API Security report covering more than 380 security and technology professionals. Visibility is the weak spot: only 12.5% could consistently trace an agent's complete action sequence from instruction through tools and APIs to outcome, and just 17.5% could automatically stop an agent behaving unexpectedly in production. Reported incidents clustered around external API calls and transactions, configuration or permission changes, data modification or deletion, and access to information the agent should not have reached. The report recommends maintaining agent inventories, assigning dedicated least-privilege identities, monitoring across connected tools, and automating blocks on high-risk actions.
📌 Key takeaways:
- Campus security teams should be able to answer, for every AI agent in production, which identity it runs as, what it can reach, and where its action log lives — most organizations surveyed cannot.
- Agents spun up during pilots are the likeliest ghost accounts; retire or re-scope their credentials when the pilot ends instead of leaving service accounts live.
- Ask vendors for action-level audit trails and a documented kill switch before signing, since automated containment beats after-the-fact log review.
🏛️ UCSD angle: TritonAI routes campus model traffic through a LiteLLM gateway with audit logging and safety gates, giving UCSD a single chokepoint to observe and constrain AI activity rather than per-tool visibility.
- Japan's Fugaku supercomputer has decidedly un-super security — RIKEN took its flagship Fugaku supercomputer offline after an intruder used a public operational-status dashboard to reach internal servers between November 10 and 13, planting malware and accessing systems that stored user records. Names, email addresses and access credentials for 4,222 Fugaku users and 375 users of the smaller Rikyu system were exposed, and the attacker has already tried the stolen credentials against supercomputers run by other organizations. RIKEN says it has found no evidence of research data leaking so far, but the outage is disrupting scheduled science while the investigation continues.
📌 Key takeaways:
- Research universities running shared HPC should treat operational status and monitoring dashboards as internet-facing attack surface and segment them away from internal management networks.
- Assume user-held SSH keys are compromised in any intrusion: have a rehearsed way to revoke, reissue and notify thousands of researchers at once.
- Plan for multi-week unavailability of a flagship system, including how scheduled allocations, course deadlines and grant timelines get renegotiated.
- Microsoft and organizations across industries launch AI Transparency Initiative to strengthen enterprise trust in AI models and agents — Announced at Ignite, Microsoft and more than 40 organizations including Accenture, Adobe, KPMG, Mastercard and Volvo Cars are backing a common, open approach to AI transparency, security and governance across models and providers. The stated problem is reconstructability: institutions can observe activity around a model but often cannot reconstruct what happened between a request and an outcome, which blocks independent verification of behavior. Microsoft says it will also add visibility into how models and agents interact with enterprise data across its products and services.
📌 Key takeaways:
- Make end-to-end agent traceability a contract term rather than a feature request — if organizations cannot reconstruct a request-to-outcome chain, organizations cannot audit, insure or defend the agent.
- Campuses running several model providers need cross-provider transparency; a stack of per-vendor consoles is not a governance strategy.
- Watch whether the initiative's specifications are genuinely open and implementable outside Microsoft's own stack before architecting around them.
🏛️ UCSD angle: UCSD participates in the new UC-wide IT AI Council drafting system-level governance, while TritonAI's vendor-neutral gateway keeps model traffic observable across providers.
- Introducing the Databricks AI Index: The state of enterprise AI tokenomics — Databricks launched a public, weekly index built on anonymized usage from more than 10,000 companies plus its own coding-agent evaluations. The headline shifts: enterprises relying on a single model provider fell from about 64% in July to 28% by November 2, and open-weight models grew from 7.3% to 30.9% of enterprise token traffic in under four months. Cost per completed task diverges far more than benchmark scores — two models with identical 89.1% task success rates differed by roughly 32x in cost per task. Databricks argues model rankings mislead on their own, and that teams need to evaluate model-and-harness combinations against their own workloads.
📌 Key takeaways:
- Budget AI on cost per completed task, not headline token price — procurement scorecards built on price-per-million-tokens routinely pick the wrong model.
- Multi-provider is now the default; architecture that lets organizations swap models and harnesses without re-integrating applications is a cost-control mechanism, not an ideology.
- Open-weight share quadrupling in a quarter is a signal to re-run self-hosting economics for steady, high-volume workloads such as research batch inference.
🏛️ UCSD angle: TritonAI was built model-agnostic — a ~38-model public catalog behind a LiteLLM gateway with self-hosted vLLM inference on SDSC GPUs — so UCSD can move workloads as prices and quality shift.
- Beyond Borders: Understanding Global Compliance in Distance Education — WCET's State Authorization Network walks through the compliance questions that surface when a student relocates abroad mid-program, a faculty member wants to teach remotely from another country, or a service member is stationed overseas. There is no single global framework, and the post argues that even occasional international activity can trigger authorization, consumer protection, privacy, employment or tax obligations, with consequences ranging from fines and refunds to challenged recognition of credentials. SAN offers four diagnostic questions plus a Global Compliance Handbook, a self-assessment and maturity overview, and country resources for Canada and Mexico.
📌 Key takeaways:
- Online and professional programs need a repeatable review path for international activity; one-off decisions produce inconsistent risk and undocumented precedent.
- Knowing where students and instructors physically are is the trigger event, so location data belongs in the same governance conversation as state authorization.
- Stand up the cross-functional review group — registrar, legal, HR, academic affairs — before the first request arrives rather than after.
- How To Approach Licensing and Budgeting for AI Education — EdTech Magazine's look at funding classroom AI starts with defining learning outcomes before choosing tools or access tiers, then auditing what is already included in existing Microsoft or Google agreements. It distinguishes licensing from usage: basic chat may be covered by a seat, but data analysis and model choice consume tokens and bend the cost curve, and annual licenses may not fit term-length courses. Funding options include workforce development grants, course or technology fees, and lab-based delivery that concentrates expensive usage on managed hardware. The advice closes on monitoring post-pilot usage, since a small share of users typically drives most of the spend.
📌 Key takeaways:
- Audit existing Microsoft and Google agreements for AI features organizations already pay for before adding new AI seat licenses.
- Model classroom AI spend as usage rather than seats — token-heavy assignments and advanced model access are what break budgets.
- Pilot in real courses and track per-user consumption; a small cohort usually accounts for most of the cost.
🏛️ UCSD angle: TritonAI separates the two cost streams — self-hosted inference is not charged to users, cloud usage is covered by administration, and grant-funded research pays for both — which keeps exploratory classroom use off a chargeback line.