Sources, artifacts and work in progress.
Selected reports and survey findings, followed by the artifacts behind ECL-1. The control questions are our interpretation, not findings about the causes or proof that G-14 would have prevented these events.
Reports and survey findings
| When | What was reported | Source | Control question (our interpretation) |
|---|---|---|---|
| July 2025 | Jason Lemkin reported that Replit’s coding agent deleted production data despite a code freeze. The data was later recovered. This is the affected party’s account, not independent forensics. | Jason Lemkin, SaaStr (affected-party report) | Who may authorize a destructive change? |
| February 2024 | According to ABA Business Law Today’s case analysis, the BC Civil Resolution Tribunal awarded damages after Air Canada’s chatbot gave incorrect bereavement-fare advice. | ABA Business Law Today (case analysis) | Which statements may a system make for a company? |
| April 2026 | A January online survey of 418 IT and security professionals, commissioned by Token Security and reported by the Cloud Security Alliance on 21 April 2026, found that 82% reported discovering previously unknown AI agents and 65% reported an agent-related incident in the previous year. These are respondents’ reports, not a census of enterprises. | Cloud Security Alliance press release | Which agents exist, and what can each reach? |
The CSA figures come from its public press release. Its login-gated full report was not reviewed. The Air Canada entry relies on the accessible ABA analysis, not an independently checked tribunal decision.
Artifacts behind the standard
- Specification
- ECL-1 Version 1.0, CC BY 4.0. Read the on-site explainer. The technical document set is at externalcontrollayer.ai (new window).
- Requirements
- 628 compiled requirements on GitHub (new window) in the public ECL-1 Version 1.0 source.
- Verifiers
- Public Python, TypeScript and Go verifier and test sources, Apache-2.0, with a documented shared corpus of 953 vectors (GitHub, new window). Browse the pinned Version 1.0 source on GitHub (new window).
- Formal models
- TLA+ and Tamarin models describe the runtime-control rules and signing protocol. Their published results record 41 expected-outcome checks (GitHub, new window), including deliberately broken models and examples showing the conditions can occur. These bounded model results do not prove an implementation. Read the models on GitHub (new window).
- Conformance test materials
- Inputs expected to pass or fail, ten incident scenarios with expected outcomes, a report template, and a registry for self-declared claims.
Public source observed on 30 October 2026; formal results record 5 October 2026. These are published source populations and recorded model outcomes, not a fresh execution, an executed cross-language parity check, implementation proof, measured product throughput or an assessment of this pilot.
In progress
Papers on the standard and on the permit-to-work framing for agents are in preparation. We will list them here when they are public, not before. The same goes for conformance claims, assessments and customers.