# Assessment finds limited public AI-model containment protocols

_Published Saturday, August 22, 2026 at 1:07 PM EDT · AI · Latest · Tier 2 — Notable_

![Assessment finds limited public AI-model containment protocols — Primary](https://techcrunch.com/wp-content/uploads/2019/07/kill-switch-ransomware.jpg?resize=1200,638)

Guidelight AI Standards assessed the public containment-response disclosures of Anthropic, Google, Meta, OpenAI and xAI, finding few published protocols for a model detected trying to subvert human control. OpenAI received 3 of 5, while Anthropic and Meta ranked lowest on public evidence. The assessment examined monitoring, response triggers, audits and plans to revoke permissions or take systems offline. The group said its results measure public disclosure, not necessarily undisclosed internal safeguards.

## Sources

- [TechCrunch](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/)

---
Canonical: https://techandbusiness.org/newswire/uDvuWg4wp-KAasOJmzBXDm
Published: 2026-08-22T17:07:19.800Z
Story chronology: 2026-08-22T16:00:00.000Z
Retrieved: 2026-10-06T23:12:47.394Z
Publisher: Tech & Business (techandbusiness.org)