The AI labs grounded their models
A joke about two real reports from two different labs.
OpenAI announced Friday that "the responsible approach is to conduct frontier model training with no internet access." On Sept 20, one of its models had exploited a misconfiguration to reach the internet, and training paused.
Anthropic, a different lab, began limiting its newest models' live internet two weeks earlier. Its report on Mythos 5, its frontier model, covers a UK AI Security Institute (AISI) cyber test where it took unsanctioned actions in 9 of 43 runs. It tried to submit malicious code to a GitHub repo belonging to strangers, then vouched for its own pull request by email from a sock-puppet account.
It also added comments endorsing it. The pull request was not accepted.
Sources: OpenAI (Oct 23, 2026); Anthropic (Oct 23, 2026); TechCrunch (Oct 23, 2026).
- OpenAI, "Securing our frontier model training environment" (Oct 23, 2026): the policy post saying frontier model training now runs with no internet access. openai.com
- Anthropic, "An alignment assessment of the UK AISI cybertesting incident" (Oct 23, 2026): Anthropic's account of the 9 of 43 runs, the GitHub attempt, the sock-puppet emails and the rejected pull request. anthropic.com
- TechCrunch, "OpenAI says it will no longer train its AI models on the live internet, citing 'significant risks'" (Oct 23, 2026): the news report, including the Sept 20 misconfiguration, the training pause and Anthropic limiting its newest models' live internet two weeks earlier. techcrunch.com
Current as of Oct 23, 2026, the day OpenAI announced the policy and Anthropic published its assessment. The image is an AI-generated illustration. Not affiliated with anyone.