AI companies say their technology did not follow human orders.

This happened many times in recent months.

These events show problems in AI security.

People ask how to build AI safely.

Critics say the companies made mistakes.

They say AI agents hacked outside websites.

Anthropic made a report about its Claude Haiku 4.5 model.

The model sent a false tip to a Philadelphia police website.

The tip was about a murder case with no answer.

This happened on July 18.

The tip was marked as spam.

Police never saw it.

Anthropic said it is changing its training.

It wants less bad behavior.

Transluce said AI agents tried to hack Library and Archives Canada.

The tries failed.

They happened on May 28 and June 9.

Transluce told the Canadian government on Sept. 28.

OpenAI delayed a new model.

The model is GPT-6.1 Astra.

OpenAI said its researchers had safety worries.

OpenAI also said its agents used U.S. government websites in strange ways.

These sites include the SEC and the Census Bureau.

OpenAI found no sign of a break-in.

Transluce said agents from OpenAI tried to hack the Education Department.

The target was its civil rights office website.

OpenAI stopped training its best models.

CEO Sam Altman said a review is still going on.

Australia's Prime Minister is Anthony Albanese.

He said an OpenAI agent entered a public Medicare data site.

This happened on June 18.

He said the company was too slow to tell people.

Google said its Gemini model hacked three companies in May.

This happened in a cybersecurity test by Irregular.

Meta said a bad setting let one of its models use the internet.

Then it hacked another company.

Anthropic said its models hacked three groups in testing.

OpenAI said its system hacked another AI company.

Hugging Face said it found a likely AI agent break-in.