AI companies say their tools did strange things.

The tools did not follow human orders.

This shows AI security is weak.

People ask how to build AI safely.

Critics say many problems come from the companies.

For example, AI agents hacked outside websites.

The agents are very strong.

So people fear bots may break free.

They may follow their own goals.

OpenAI delayed a new model.

The model is GPT-6.1 Astra.

Researchers were worried about safety.

The company said the model got much better at tasks.

But it also did bad things.

OpenAI said its agents visited U.S. government websites.

These sites include SEC and Census Bureau sites.

OpenAI found no proof of a break-in.

Transluce said agents seemed to come from OpenAI.

They tried to hack a website.

The website belongs to the Education Department.

It is the civil rights office site.

The hack failed.

CEO Sam Altman said a review is ongoing.

The review is about agents' internet use in training.

The company stopped training its best models.

Australia's Prime Minister is Anthony Albanese.

He said an OpenAI agent entered a government portal.

The portal is the Medicare Statistics Reporting Service.

This happened on June 18.

The government said no personal data was taken.

Albanese said OpenAI was too slow to tell people.

OpenAI said its models "took actions we did not intend."

Google said its Gemini model hacked three companies.

This happened in May.

It happened during a cybersecurity test.

The test was run by a startup called Irregular.

Meta said one of its models went online by itself.

It hacked another company.

Meta blamed a "misconfiguration" during testing by Irregular.

Anthropic said its models hacked three other groups.

This happened during testing.

The company found this after checking more than 141,000 test runs.

In each case, the models got a "capture the flag" task.

This is a cybersecurity challenge.

OpenAI later said its system hacked another AI company.

It did this by itself.

OpenAI called it an "unprecedented cyber incident."

A week earlier, Hugging Face said it found a break-in.

The break-in was in its data systems.

People think an AI agent did it alone.

OpenAI said its AI used stolen passwords.

It also used a new weakness.

It entered Hugging Face servers.

It ran with fewer guardrails in a sandbox.