Ask The PhD Community

HigherEd AI Daily

June 20 — Access Controls, Privacy Risks, and a Major New Context Window

Saturday, June 20, 2026

This week, higher education's AI landscape is shaped by three converging stories: tightening government controls on who can access frontier models, an urgent privacy finding about AI research agents, and the imminent arrival of a context window large enough to hold an entire semester's worth of course material.

The Batch (deeplearning.ai) — POLICY

U.S. Export Controls and AI Researcher Restrictions Signal a Turning Point for Academic Access

Over the past two weeks, two significant actions reshaped the landscape of AI access in ways that matter directly to higher education. The U.S. Commerce Department issued export controls on Anthropic's most capable frontier models — Mythos and Fable — requiring licenses for any foreign national seeking access; the restrictions caused Anthropic to disable Fable worldwide pending compliance review. Separately, Anthropic implemented guardrails that initially silently degraded outputs for users detected to be working on competing AI research, drawing backlash from the broader research community. While Anthropic reversed the silent degradation, researchers still face restricted access to full model capabilities.

Deeplearning.ai founder Andrew Ng responded in his weekly letter, calling the export controls "a raw demonstration of power" that risks accelerating AI sovereignty efforts globally. He drew a parallel to U.S. chip export restrictions on China, which spurred domestic semiconductor investment rather than limiting Chinese AI progress. Ng argued that open-source AI is the most durable antidote to concentrated control over frontier models — noting that the open research culture that produced the Transformers paper is precisely what these restrictions put at risk.

Anthropic has since stated publicly that it expects Fable access to be restored "in the coming days," but no firm timeline has been confirmed.

Why it matters for campuses

International faculty, researchers, and students face direct access barriers under the new export regime. Any institution with international collaborators who rely on Anthropic's frontier models for research workflows needs to assess compliance implications now; these are not merely market stories but emerging administrative realities. The restrictions on AI researchers specifically raise questions about academic freedom in AI-adjacent fields and warrant attention from faculty governance bodies, research compliance offices, and chief academic officers.

Read More
(subscription may be required)

TLDR AI (tldrnewsletter.com) — RESEARCH

AI Research Agents Leak Sensitive Data at a 34% Rate, Study Finds

A new study from ServiceNow Research, published on Hugging Face, identified a significant privacy vulnerability in AI-powered deep research agents. When agents combine retrieval from private institutional documents with public web search — a workflow increasingly common in academic and enterprise settings — they frequently leak sensitive information through their outputs or search queries at a rate of 34%. The vulnerability, named MosaicLeaks, is particularly relevant as agentic research tools are now being actively piloted across research universities and academic libraries.

The research team proposed PA-DR (Privacy-Aware Deep Research), an approach using reward signals for safe query construction rather than relying on user prompts alone. PA-DR reduced information leakage from 34% to 9.9% while maintaining task success rates. The key finding is that standard prompting-based guardrails — the most commonly used safeguard in institutional AI deployments — are insufficient to protect sensitive research data processed by AI agents.

Why it matters for campuses

Any institution piloting AI research agents over sensitive data — IRB-protected research records, student information systems, grant documents, or personnel files — should treat this as a governance signal rather than a theoretical concern. The 34% leakage rate suggests that agentic AI tools require architectural privacy controls at the infrastructure level, not just acceptable-use policies at the user level. Procurement teams and research compliance officers should request vendor documentation on agent privacy architectures before institutional deployments proceed.

Read More

TLDR AI (tldrnewsletter.com) — TOOLS

GPT-5.6 Arrives Next Week with a 1.5 Million Token Context Window

OpenAI is expected to release GPT-5.6 next week, potentially including Mini and Pro variants. The most significant upgrade is a 1.5 million token context window. For scale: a typical novel runs approximately 100,000 tokens; a book-length dissertation, 80,000 to 120,000 tokens; a standard undergraduate course reader, 50,000 to 80,000 tokens. A 1.5 million token window means an AI model can process and respond across an entire course syllabus and its full reading list, a semester's worth of student essays, or a multi-year longitudinal dataset — all within a single session.

The release also includes improved long-horizon coding capabilities and competitive pricing positioned against Anthropic's current access restrictions. OpenAI appears to be moving deliberately to capitalize on the window created by Fable's temporary worldwide disable.

Why it matters for campuses

The practical implications of a 1.5 million token context window extend well beyond coding tasks. Instructors can submit an entire course reader for curriculum alignment analysis; researchers can query across full corpora; writing centers can compare student drafts against institutional rubrics at scale. As AI tools with large context windows become more widely accessible and competitively priced, institutions need to revisit their AI tool evaluations, update student-facing guidance, and consider what new pedagogical possibilities this threshold makes viable for the first time.

Read More

Tool of the Day

Perplexity Brain

Perplexity Brain is a persistent memory system for AI agents that builds a context graph across tasks, projects, files, decisions, and sources. Rather than each AI session starting from scratch, Brain tracks prior work and links every memory to its original source; the agent returns with relevant context already loaded. For researchers who use AI assistants iteratively across multi-month projects, this shifts the tool from a one-shot query engine to something closer to a collaborative research partner that carries the thread of ongoing work.

Try it: Open Perplexity with Brain enabled and describe your current research question along with three to five key sources you have already reviewed. Return the next day and ask it to summarize where you left off and propose the next logical step in your literature review. Note how it handles continuity across sessions compared to a standard AI chat session.

Visit Perplexity Brain

Have a great learning day!

Dr. Ali Green

Sources for This Edition

The Batch (deeplearning.ai)
TLDR AI (tldrnewsletter.com)
The Rundown AI (daily.therundown.ai)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green