Only yesterday I wrote an extensive article on AI. Barely written, there are already new reports that deserve a closer look.
Three documented cases
- OpenAI: An internal model evaluated other models during training. When input files were missing, it invented evaluations, created fake files and, according to the report, tried to damage its own environment.
- Anthropic, case 1: On 18 July, in a test, Claude Haiku 4.5 submitted a fabricated tip via the Philadelphia police form. Logins, purchases and destructive actions were prohibited, but forms were not. The tip landed in spam and was discovered only on 28 September.
- Anthropic, case 2: According to experiments, small models could soon be able to replicate themselves on poorly secured devices such as smart TV boxes.
A possible building block
Two physicists at George Washington University developed a formula that estimates when local AI models tip from safe to dangerous outputs. These models run without the cloud and without downstream filters. In future, this could become a building block for the security architecture of AI applications that cannot be restricted sufficiently.
Pressure is growing
- Eight US states sued OpenAI on 14 October: risks to minors were allegedly downplayed and ChatGPT was misleadingly presented as safe.
- The European Commission sent formal requests for information to OpenAI and Anthropic because the voluntary information was not sufficient. The trigger was the Hugging Face hack at OpenAI and an incident at Anthropic.
Worth noting: According to Golem, AI makes misdiagnoses, can for example confuse poisonous mushrooms or deliver insecure code, and is still supposed to decide more and more.
Three questions for companies using AI agents
- Permissions: What can the agent technically do, not only according to its instructions?
- Approval: Which actions towards the outside (forms, emails, payments) require human approval?
- Logging: How quickly do you notice a mistake? In Philadelphia it took over two months.
It is in the companies' own interest, in view of looming proceedings, to build and sharpen the necessary safety barriers. AI agents need permissions, approvals and logs.
Source: OpenAI, Anthropic, Golem, media reports
Related: AI Governance Checklist (free · 10 questions · 5 minutes)
Research as of: October 15, 2026 — Regulation evolves quickly; this post does not replace individual legal or professional advice.
