- AI has made it ‘close to impossible’ to assess university coursework, regulator says — England’s Office for Students investigated AI use in essays, dissertations and “open book” exams and concluded that the credibility of English degrees is under threat, with students reporting that peers boast of using AI undetected by staff. The regulator warned universities that failing to design “valid and reliable measures of students’ knowledge and skills” could breach their conditions of registration and lead to sanctions. Its report, produced with Advance HE from interviews with more than 300 students and staff at 127 universities and colleges, points to greater use of in-person written exams, oral vivas, process-based assessment and exercises where students critique or edit AI-generated output. Josh Fleming, the OfS director of strategy and delivery, called AI use “a generational risk to the system as it stands” and asked for “far greater use of AI-free assessment.”
📌 Key takeaways:
- Assessment design has become a compliance question, not just a pedagogical one: regulators are beginning to tie AI-resilient assessment to registration conditions, so provosts and CIOs should expect evidence requests about how programs measure student knowledge, not just advisory guidance.
- The practical menu is already established — in-person exams, vivas, process documentation, controlled-AI exercises — which means assessment committees can stop debating whether to change and start sequencing changes by program risk.
- Programs built almost entirely on take-home work face the largest exposure, and the fix is expensive in faculty time; budget instructional-design support alongside any policy announcement, or the policy will not survive the first semester.
🏛️ UCSD angle: Rather than buying proctoring software, UCSD has been framing this as an identity problem: a TritonAI pipeline fuses Canvas submission telemetry with LiteLLM gateway prompt logs to flag agent-completed coursework without cameras or lockdown browsers.
- New AI Learning Matrix Brings Nuance to ‘Wild West’ Moment — Researchers at Harvard’s Compass Collaborative drew on more than 5,000 human development, education and psychology studies to build the AI Guidance by Age matrix, launched this week for educators, caregivers and learners. Users select a role from nine options, a developmental domain such as cognitive foundations, social understanding or language and literacy, and an age range; the tool returns recommended uses, assessment-design advice, and where AI is likely to help or hinder a specific learning goal. Co-creator Jeffrey James Curley told Inside Higher Ed that many campus decisions have “centered on technology and not on learner needs,” and that higher education has been slow to acknowledge students are under pressure to use AI in ways with short-term value and long-term cost. The article notes one survey finding that 85 percent of college students use large language models for brainstorming, research, writing and editing, while permissibility still varies course by course at most institutions.
📌 Key takeaways:
- A developmental lens gives curriculum committees something more defensible than blanket permission or blanket bans, because the same tool use can legitimately be appropriate for an upper-division seminar and harmful for a student still building foundational skills.
- Course-by-course discretion is the real governance gap: when AI rules live entirely with individual instructors, students experience inconsistent expectations across departments, and a shared research-backed framework is the cheapest way to reduce that variance.
- Much of the underlying evidence predates generative AI, so campuses should treat the matrix as a structured starting point for local data collection rather than a settled answer to hand faculty.
🏛️ UCSD angle: UCSD’s AI Readiness Navigator gives faculty a self-assessment that returns a personalized upskilling path into the governed tool ecosystem, and the proposed Course Redesign Studio would push that same guidance down to individual assignment and syllabus design.
- Closing the AI guidance gap in education — OpenAI published a Responsible AI in Education report drawing on new ThinkYoung research that surveyed 5,000 secondary students aged 13–19 across EMEA, alongside lessons from education partnerships in countries including Estonia, Slovakia and Greece. The central finding is a gap between use and instruction: 79 percent of surveyed teenagers had used AI for homework or assignments, but only 47 percent said anyone had shown them how. The report argues that teachers become more important rather than less as AI access spreads, and organizes its recommendations around equitable access to frontier tools, teacher training and support, and continuous evaluation of what actually improves learning.
📌 Key takeaways:
- The adoption-versus-guidance gap is the operational problem on most campuses: usage is already near-universal while formal instruction in how to use the tools lags far behind, which makes faculty and staff development the binding constraint rather than licensing or model access.
- Vendor-authored education guidance is arriving quickly and will shape procurement conversations; technology leaders should read it as an input to local policy and evidence gathering, not as a substitute for institutional governance.
- The 13–19 age band matters for dual-enrollment, early-college and pre-college programs, where K–12 rules and university AI policy meet and frequently conflict.
🏛️ UCSD angle: TritonAI’s own experience matches the finding — broad TritonGPT adoption without a structured path from awareness to fluency is why UCSD scoped the AI Readiness Navigator and an AI Literacy Micro-Credential as a sequenced program rather than a single training course.
- Cyber Criminals Move Toward More Targeted Phishing Attacks — KnowBe4’s Phishing Threat Trends Report, Vol. 8, built on roughly 1.15 billion daily event signals, found overall phishing volume up 29.8 percent year over year while spear phishing climbed 41.1 percent and company impersonation rose 28 percent. Attackers are moving away from mass campaigns toward researched, personalized messages that borrow the identity of coworkers, businesses and technology platforms; business email compromise accounted for 51.2 percent of analyzed messages, with HR, finance and executive leadership the most frequently targeted groups. The report also found that 41.3 percent of dangerous emails bypassed secure email gateways, and that phishing was twice as likely to originate from legitimate infrastructure in Q3 2026 as in the previous quarter — including AI services used to host malicious content on trusted domains and abuse of OAuth permissions. Nearly all messages, 90.4 percent, arrived Monday through Friday, with Monday the single busiest day at 20.6 percent.
📌 Key takeaways:
- Mail filtering alone is no longer a defensible control: when roughly four in ten dangerous messages clear secure email gateways and attackers rent trusted infrastructure, campus security teams need identity-layer defenses — phishing-resistant MFA, OAuth consent governance and conditional access — working alongside the gateway.
- OAuth permission abuse and legitimate-platform hosting should be the focus of this year’s tabletop exercises, since awareness training built around typos and fake IT support does not address a message that arrives from a real service the campus already uses.
- The weekday concentration is a staffing signal for the SOC and a timing cue for simulated phishing and awareness pushes, which land harder early in the week.
🏛️ UCSD angle: The shipped Phishing Analyzer covers email-borne threats, which is the gap the TritonAI SecOps Alert Triage Agent was scoped against — clustering, enriching and suppressing false positives across the Splunk and OpenSearch monitoring stacks so analysts stop chasing duplicates.
- Technology Enables Higher Ed IT Leaders to Make More Proactive Decisions — EdTech Magazine describes how campuses are shifting from improvised, reactive technology decisions toward intentional ones as AI spreads across campus, cyberthreats intensify and classroom technology becomes more central to teaching. Institutions are standing up cross-departmental task forces to shape AI policy and governance so practices stay consistent and IT keeps visibility into which tools are actually part of the campus ecosystem. That visibility is extending to third-party providers, with security leaders building assessments vendors must meet and procedures for responding when a breach occurs. Classrooms are being standardized so faculty encounter the same technology in every room, and pandemic-era device loaner programs have evolved into year-round services for students in need.
📌 Key takeaways:
- The governance lever here is visibility rather than prohibition: cross-departmental AI task forces only work if IT can enumerate the tools actually in use, which argues for putting procurement, SSO and a model gateway in front of pilots instead of chasing a shadow-AI survey after the fact.
- Vendor risk assessment belongs inside the AI onboarding path rather than after contract signature — the questionnaire that covers a data breach should also cover model training, data retention and subprocessors.
- Classroom standardization and device loaner programs look unglamorous next to AI pilots, but they are the operational base that makes AI-enabled instruction reliable across a heterogeneous campus.
🏛️ UCSD angle: Centralizing model traffic through the TritonAI LiteLLM gateway is how UCSD gets that enumeration — every application and agent using campus-sanctioned models is attributable by SSO, so the inventory builds itself instead of arriving as an annual survey.
- Introducing AlphaProof Code: a breakthrough in verifying complex software at depth — Google DeepMind researchers used a multi-LLM agentic framework called AlphaProof Code to formally prove key security properties of the actual code underlying the Signal protocol — the first time AI has formally verified a widely deployed cryptographic protocol at this depth, and the first time such an implementation has been verified by any method. Human experts worked with models to translate Signal’s security goals into formal specifications, then agents powered by Gemini Carbon generated more than six million lines of Lean proving the compiled code meets them down to the code running on devices. The agents audited their own proofs and flagged about one percent of the codebase for human review, finishing in roughly four weeks work that historically takes expert teams years. Reintroducing two previously fixed bugs produced machine-checked proofs that the buggy versions were insecure.
📌 Key takeaways:
- Formal verification is moving toward procurement vocabulary: research universities and security teams should expect vendors to begin claiming machine-checked proofs, and should ask what precisely was specified, by whom, and which checker validated it.
- The division of labor — humans write the specification, agents generate proofs, a deterministic checker validates every one — is the strongest available pattern for governing agentic AI, because correctness is machine-verifiable rather than dependent on reviewing every artifact a model produces.
- Four weeks versus years changes the economics of assurance for cryptographic and safety-critical code that campuses depend on but rarely audit, from authentication libraries to research instruments.