CCC ISTI 25-300-02 AI Safety and Security
Citation
Hayward, Craig, Vice Chancellor, AI Strategy and Digital Transformation. “AI Memo Series, Part 2 - Centering AI Safety and Security Across the California Community Colleges.” Memorandum ISTI 25-300-02. California Community Colleges Chancellor’s Office, May 14, 2026. https://www.cccco.edu/wp-content/uploads/2026/08/ai-memo-safety-and-security-may-2026-a11y.pdf. Copy: raw/shared/2026-10-13 CCC ISTI 25-300-02 AI Safety and Security.md (original: .pdf beside it), fetched 2026-10-13. Read 2026-10-15.
Bottom line
- The California Community Colleges (CCC) Chancellor’s Office tells colleges to test agentic AI in isolated sandboxes. It also asks for a human in the loop and vetted enterprise tools for student data.
- The memo names six enterprise products that it says protect institutional data. This is a list of the AI vendors already in the CCC system.
- It also shows a demand for safe test environments that colleges must now build alone.
Background
This is the second memo in the CCC spring 2026 AI guidance series. It builds on ISTI 25-300-01 (AI literacy). It asks colleges to review the guidance before the end of the spring 2026 term.
Key claims
- Scenario A: AI bots for student support need “Human-in-the-Loop (HITL) protocols” and a tested “escalation path” to a human. (p. 2)
- Scenario B: “it is best practice to evaluate agentic tools in a sandbox environment before broader deployment.” Technical leaders should use “isolated physical or virtual environments and dedicated ‘low-privilege’ API keys” and test for prompt injection. (pp. 2 to 3)
- Suggested HITL procedures: human authorization, identity mapping of agent actions to a human, limited access, boundary setting and disclosure. (p. 3)
- Scenario C: faculty and students should use tools “under institutional or enterprise licenses, such as Nectir and Playlab, that are vetted and compliant with FERPA”. (pp. 3 to 4)
- For deployers: “prioritize enterprise subscriptions (e.g., Google Workspace for Education, ChatGPT Edu, Microsoft 365 Copilot, Claude Enterprise Plan, Nectir, Playlab)” that keep data out of model training. (p. 4)
- Before connecting a tool to Canvas, deployers should confirm it meets “WCAG 2.2 Level AA accessibility and user privacy standards”. (p. 4)
- Leadership should define permitted use “through tiered models, such as Open, Conditional, Restricted, or Closed”. (p. 5)
- Leadership should align security with “the cybersecurity allocation provided in AB 178” and the CCC Security Center. (p. 5)
Evidence and method
Guidance with example scenarios. It names one pilot (“the Cali financial aid pilot”) but gives no results.
Limits
- Recommendations only. No data on how many colleges follow them.
- The product list is a set of examples, not a procurement decision.
- The memo names FERPA (the Family Educational Rights and Privacy Act) and SOC 2, but it gives no test for compliance.