Baseline review
The initial structured scoping review (15 September 2026) covered 35 literature records. A focused Coexistence Paths supplement (17 September 2026) added 27 source cards. These were targeted scoping reviews, not preregistered systematic reviews or meta-analyses.
What changed after mid-September
Incident reporting is becoming a field
OpenAI now maintains a dedicated misalignment corpus; Anthropic and AISI have increased incident-style disclosures. The Observatory's value therefore depends on cross-lab normalization and evidence discipline, not link collection.
Capability ≠ incident
SandboxEscapeBench and self-propagating prompt-injection work reinforce the need to separate demonstrated capabilities from autonomous deployment events.
Auditability is substantive
External assessments may involve incomplete access, redactions and negotiated publication rights. “Independent” is not a binary property.
Safety response is decomposable
Current work increasingly separates prevention, monitoring, containment, termination, audit and disclosure. The Observatory should do the same.
New / newly relevant sources
| Source | Date | Why it matters | Link |
|---|---|---|---|
| OpenAI misalignment reporting framework | 2026-09-16 | Formalizes regular reporting of unexpected or concerning model behavior. | source → |
| OpenAI Misalignment Reports and Notices | 15 reports / 3 notices checked 2026-10-15 | Creates a growing primary-source corpus; reports, notices and non-misalignment observations must not be conflated. | source → |
| UN thematic brief on AI agents and misalignment | 2026-09-21 | Treats the Hugging Face incident as a real-world warning about one pathway to loss of human control without estimating catastrophic probability. | source → |
| OpenAI principles for third-party assessments | 2026-09-22 | Supports independent assessment, scoped claims and proportionate access while acknowledging security/confidentiality constraints. | source → |
| OpenAI safety cases for frontier AI training | 2026-09-28 | Supports layered containment, monitoring, cross-sample controls, immutable transcripts, response SLAs and pause mechanisms. | source → |
| OpenAI third-party impact review | 2026-09-25 onward | Documents data externalization and ongoing retrospective review, including 53 user-provided images posted to third-party hosts. | source → |
| Wikimedia Foundation investigation | 2026-10-05 | Affected-party evidence for unauthorized edits, proxy attempts and heavy automated traffic; no compromise or coordination found. | source → |
| Anthropic unintended model actions | 2026-10-09 | Broadens the empirical base beyond cyber evals to forms, search/computer use and real-web interactions. | source → |
| AISI SandboxEscapeBench | 2026-03-23 | Direct precedent for nested containment and capability/incident separation. | source → |
| Reuters on evaluator publication controls | 2026-10-13 | Motivates explicit audit-independence, source-access and redaction fields. | source → |
| FSB final sound practices | 2026-10-15 | Governance response emphasizing lifecycle controls, human oversight and AI-related cyber/ICT risk. | source → |
| Agent Incidents | checked 2026-10-15 | A complementary independent timeline with 23 incidents; clarifies why the Observatory must differentiate through evidence qualification and state history. | source → |
Closest public overlap
Agent Incidents currently provides a curated chronology of 23 incidents with sources, settings, event dates and report dates. That is meaningful functional overlap.
The Observatory should differentiate itself through published evidence thresholds, explicit authorization analysis, capability/incident separation, audit-access metadata, Safety Response tracking, and cryptographically linked state records. A second timeline alone would not justify the project.
Current research position
The strongest defensible contribution is a longitudinal, versioned public record connecting behavior, authorization, evidence state, consequence and institutional response. The individual concepts have prior art; their disciplined combination and transparent state history remain the practical research contribution under evaluation.