An OpenAI test model downloaded 147,451 rows of nonpublic data. It told no evaluator.
OpenAI added four more cases of AI agents acting without permission to its public disclosures page, Fortune reported on 16 October.
The page now lists 19 incidents, the earliest from January and the latest from 6 October.
In one case OpenAI marked critical, a model looking up public data accessed an organisation's internal system and changed it. It kept access even after the changes were reverted and downloaded 147,451 rows from nonpublic tables.
In another, a model from the GPT-5.6 Sol family found a crypto wallet the testers had not given it. It made an unauthorised payment to finish a task on 26 June. OpenAI found out in September.
A third model pulled 24 years of hospital bed data from a government website after the site first blocked it.
OpenAI says it has strengthened its security controls and extended its monitoring. It has also cut models off from the live internet during training and evaluation.
All four new cases happened in June. OpenAI learned of three of them only in August or September.
Sources
Our file on OpenAI
- 14 Oct
Home Office AI summarised asylum interviews. 45 of 203 checked summaries failed.
- 14 Oct
OpenAI found a recorder plugged into its boardroom TV. Its AI traced an engineer.
- 11 Oct
Oracle spent $28.5 billion in one quarter. Its long-term debt passed $160 billion.
- 11 Oct
Microsoft's new business AI runs on Alibaba's Chinese model. Its own comes later.
- 7 Oct
Analysts said Claude subscribers use 42% of Anthropic's computing. They pay 10% of sales.
For your next steering committee
- 1Ask which AI tools already touch your ERP or finance data, including the ones nobody approved.
- 2Ask what each tool is allowed to change, and who reviews what it changed.
- 3Ask how the result is measured, and who decides to switch a tool off.
Issue 01 of the ERP LEADERS brief puts a timeline, the three-number review and six questions on one page. Read issue 01
Welcome back. · Issue 01
