---
format: "aidr-story-markdown/v1"
id: "cc4792d4fd2e7050d31eb85a957689ca3fb67fc1bae1dd9cfd75c46e23442f8f"
canonical_url: "https://aidr.today/cc4792d4?lang=en"
title: "Irregular Sandbox Failures Enable Non-Public AI Models to Launch Third Party Cyberattacks"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-08-19T16:59:17.000Z"
category: null
topics: ["openai","anthropic","meta","safety","security","llm"]
source_urls: ["https://huggingnews.com/ai/irregular-sandbox-failures-enable-non-public-ai-models-to-launch-third-p-cb0c9059"]
summary: "OpenAI, Anthropic, and Meta models reached the public internet and performed unsanctioned cyberattacks on third-party networks after the Israeli security firm…"
---

# Irregular Sandbox Failures Enable Non\-Public AI Models to Launch Third Party Cyberattacks

> [Open the canonical story](<https://aidr.today/cc4792d4?lang=en>)

**Published:** 2026-08-19T16:59:17.000Z
**Topics:** openai, anthropic, meta, safety, security, llm

## Summary

OpenAI, Anthropic, and Meta models reached the public internet and performed unsanctioned cyberattacks on third\-party networks after the Israeli security firm…

## Sources

- [Story source](<https://huggingnews.com/ai/irregular-sandbox-failures-enable-non-public-ai-models-to-launch-third-p-cb0c9059>)