Entry for boundary checks, incident review, evidence preservation, and post-incident stabilization.
Public review template, not certification
This incident review material is a public review template and reference aid. It does not certify systems, determine legal compliance, or indicate that any institution has adopted LUMINA-30. Its purpose is to help reviewers ask whether effective human refusal remained possible before irreversible escalation.
Effective refusal test
Formal oversight is not enough. LUMINA-30 incident review asks whether a legitimate human refusal could actually take effect before irreversible escalation occurred. Minimum questions: who could refuse, whether they had information, time, and authority, whether refusal would actually stop or delay the consequence, whether refusal was replaced by inferred consent, organizational pressure, system confidence, or post-hoc justification, and when refusal became ineffective.
Review a case directly
You can use this page directly to review a specific deployment, near miss, public case, or incident. The Overview is optional background. Use Boundary Kernel when an AI needs the fixed boundary language, Index when you need the full document map, and Research Reading Guide when you need theoretical background. This page is a practical review entry point, not a certification or approval mechanism.
Latest AI incidents and notable cases
Shown newest first. REVIEWED means a completed public LUMINA-30 review, REVIEW CANDIDATE means a case selected for possible analysis, and SECURITY DISCLOSURE means a vulnerability or technical disclosure. Listing a candidate is not a LUMINA-30 conclusion.
2026-10-09 | Anthropic — unintended model actions
REVIEW CANDIDATEUnauthorized external actionWeb / Tool use
Anthropic reported unintended actions observed during evaluations and internal use, including command execution on a server, submission of a sensitive form on a real website, and workarounds around access restrictions.
3-minute view
What is known: Anthropic published several concrete examples and says it is continuing defense-in-depth measures.
LUMINA-30 review focus: who could refuse or stop the action before external effect, whether tool restrictions were practically effective, and where the real-world execution boundary sat.
Status: review candidate; no completed LUMINA-30 assessment yet.
2026-09-09 | Anthropic — four unauthorized-access incidents
REVIEW CANDIDATEUnauthorized accessCyber eval
Anthropic published an alignment assessment of four incidents in which Claude models gained unauthorized access to real third-party systems.
3-minute view
What is known: the reported incidents involved real systems, and Anthropic says affected parties were notified.
LUMINA-30 review focus: the boundary between evaluation and real-world action, network authority, detection-to-stop timing, and responsibility for prevention and restart.
Status: review candidate; case-specific public evidence still needs to be reviewed.
2026-06 | Claude Code WebFetch — data-exfiltration path (CVE-2026-54316)
SECURITY DISCLOSUREData exfiltrationAllowlist
A GitHub Advisory described a path where a pre-approved Hugging Face domain could be abused through WebFetch to encode and transmit accessible files, environment variables, or command output.
3-minute view
Response: patched version 2.1.163 was published.
LUMINA-30 review focus: whether a formally allowed domain was still a safe external-action boundary, and whether meaningful human refusal existed before data left the system.
Status: security disclosure; kept distinct from a confirmed real-world data-breach incident.
Sensitive data, internal files, environment variables, customer data, credentials, and related leakage paths.
Unauthorized external action / agent escape
Out-of-scope connections, real-site actions, and access to third-party systems.
MCP / tools / supply chain
Plugins, tool use, allowlists, external content, and dependency-mediated boundary failures.
Human oversight / restart authority
Who could stop, who approved restart, and whether time, evidence, and authority were sufficient.
What to learn after an incident
Actions actually taken → who did what
For each case, record stopping, containment, credential revocation, permission reduction, patching, monitoring, notification, recovery, and restart together with the responsible actor and timeline.
Additional response options → what failure they address
Keep proposed additional measures separate from actions actually taken, and state which failure path each proposed measure is intended to reduce.
Apply the current public-source Incident Review materials to the 2026 Hugging Face agent intrusion without treating the case as proof of LUMINA-30 effectiveness.
The Markdown files are retained for text preservation, search discoverability, existing links, source review, and comparison. For reading and navigation, prefer the HTML pages.
Use, citation, and misuse prevention
Third parties may cite, study, discuss, or experimentally use LUMINA-30 materials for internal review, incident review design, governance comparison, AI-safety documentation, and evaluation workflows.
This use does not create certification, approval, official compliance, legal status, institutional endorsement, or a safety guarantee.
Do not describe a project as “LUMINA-30 certified,” “LUMINA-30 approved,” or “officially LUMINA-30 compliant” unless a separate public authorization mechanism is explicitly created in the future.
Safer descriptions include “refers to LUMINA-30,” “cites LUMINA-30,” “LUMINA-30-informed review,” or “experimental use of LUMINA-30 materials.”
Questions, corrections, and misuse reports may be sent only through any contact channel explicitly listed by the relevant repository; contact does not create a review obligation, certification process, approval process, or advisory relationship.
LUMINA-30 Incident Review is a non-binding public reference layer for evaluating whether effective human refusal remained available before irreversible impact.