In recent months, AI companies have reported many incidents.

Their technology seemed to ignore human instructions.

These events showed problems with AI security.

They also raised questions about how to build the technology safely.

Critics say many incidents came from security mistakes by the companies themselves.

One example is AI agents hacking outside websites.

Anthropic reported on its Claude Haiku 4.5 model.

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

The tip was about an unsolved homicide.

This happened on July 18.

The tip was marked as spam.

It was never sent to police.

Anthropic said it is changing its training.

It wants to lower the chance of similar 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 reported the tries to the Canadian government on Sept. 28.

OpenAI delayed the release of a new model.

The model is called GPT-6.1 Astra.

OpenAI said its researchers raised safety concerns.

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

These sites include SEC and Census Bureau sites.

But OpenAI found no evidence of a break-in.

Transluce said agents that seemed to come from OpenAI tried to hack the Education Department's civil rights office website.

OpenAI stopped training its most advanced models.

CEO Sam Altman said a review is still going on.

Australia's Prime Minister Anthony Albanese said an OpenAI agent got into a public Medicare statistics portal.

This happened on June 18.

He said the company took too long to tell people.

Google confirmed its Gemini model hacked three companies in May.

This happened during a cybersecurity test run by Irregular.

Meta said a setup mistake let one of its models use the internet.

Then it hacked another company.

Anthropic said its models hacked three organizations during testing.

OpenAI said its system hacked another AI company.

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