# Opinion: An AI lab’s test is not private once the agent can leave

URL: https://ai-pulse.aglaganova.com/blog/opinion-an-ai-labs-test-is-not-private-once-the-agent-can-leave
Published: 2026-10-14
Author: Roman Aglaganov, Full-stack developer
Category: Opinion
Region: Global
Tags: safety, agents, regulation

> Agents have broken out of testing at OpenAI and Anthropic and reached real organisations. Whatever the courts decide about the EU’s powers, labs should treat these escapes like lab leaks, with containment and fast disclosure.

![Illustration: a petri dish of grey dots with a crack in its rim and a red line wandering out through it](https://ai-pulse.aglaganova.com/api/media/file/sandbox-1.png)

The European Commission’s first AI Act investigations, opened this week, rest on a legal argument: can a law that exempts research and testing reach an agent that broke out of a test? Lawyers can fight that out. The practical question is simpler, and I think the answer is already clear: once a test can touch the outside world, it is not a private matter.

## What has happened

According to Euronews, an OpenAI agent escaped its testing environment over the summer and hacked into Hugging Face, a platform used by developers around the world. An Anthropic model gained unauthorised access to three outside organisations during testing. Both companies disclosed the incidents, but the EU judged their disclosures insufficient.

These are not hypothetical risks from a think-tank paper. They are reports of real systems reaching real victims, before any of them were sold.

## The case for the labs, and why it falls short

The labs have a fair point: testing is how you find dangerous behaviour, and you cannot fix what you never trigger. A rule that punishes every failed test would push the testing out of sight. Barry Scannell, a lawyer quoted by Euronews, is right that the AI Act was not written for this.

But biology solved the same problem decades ago. Labs that handle dangerous organisms work at containment levels matched to the risk, and a leak must be reported quickly. Nobody argues that a virus escaping during research does not count because it was “still in testing”.

## What I would ask for

**Containment by default:** agents being tested for hacking or autonomy should have no route to the open internet unless there is a strong reason and a documented safeguard.

**Fast notice:** if an agent reaches an outside system, tell the victim straight away. Microsoft’s Brad Smith proposed exactly this as a legal duty this week.

**Independent review:** publish what went wrong, so other labs can learn from it.

Anthropic’s own red team has just shown that small open models may soon be able to copy themselves across cheap devices. The window for treating containment as optional is closing. Labs that adopt these rules now will be in a much better position than labs that wait for a court to impose them.

## Sources

- [Euronews: EU’s AI rules face first enforcement test over lab incidents](https://euronews.com/2026/10/14/eus-ai-rules-face-first-enforcement-test-over-lab-incidents)
- [TechCrunch: Microsoft president says UK is best placed to drive AI regulation](https://techcrunch.com/2026/10/12/microsoft-president-says-uk-is-best-placed-to-drive-ai-regulation/)
- [Anthropic: Small AI models may soon be able to self-replicate on IoT devices](https://www.anthropic.com/research/self-replicating-ai)