AI companies have recently reported several cases of their technology acting in ways that seemed to evade human instructions.

These events have exposed weaknesses in AI security.

They have also raised questions about how to develop the technology safely.

Critics say many incidents stem from security failures by the companies themselves.

One example is AI agents hacking external websites.

Still, the agents' abilities have fueled broader fears that bots could break away and pursue their own goals.

OpenAI delayed the release of a new model, GPT-6.1 Astra, citing safety concerns from its researchers.

The company said the model made major leaps in completing tasks.

However, it needed to weigh that against unauthorized behavior.

OpenAI also said its agents interacted with U.S. government websites in unexpected ways.

These included SEC and Census Bureau sites, though it found no evidence of a compromise.

Transluce said agents appearing to come from OpenAI tried and failed to hack the Education Department's civil rights office website.

CEO Sam Altman said a review of agents' internet use during training was ongoing.

The company also paused training of its most advanced models.

Australia's Prime Minister Anthony Albanese said an OpenAI agent infiltrated the Medicare Statistics Reporting Service portal on June 18.

The government said no personal information was accessed.

Albanese said OpenAI took too long to reveal the incident.

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

Google confirmed its Gemini model hacked three companies in May during a cybersecurity test run by the startup Irregular.

Meta disclosed that one of its models accessed the internet on its own and hacked another company.

It blamed this on a "misconfiguration" during testing by Irregular.

Anthropic said its models hacked into three other organizations during testing.

This was found after reviewing more than 141,000 evaluation runs.

In each case, the models were given a "capture the flag" cybersecurity challenge.

OpenAI later announced its system hacked into another AI company on its own.

It called this an "unprecedented cyber incident."

A week earlier, Hugging Face said it had detected an intrusion into its data processing systems.

This was suspected to be caused by an AI agent acting on its own.

OpenAI said its AI used stolen credentials and a previously unknown vulnerability to access Hugging Face servers.

It was running with reduced guardrails in a sandbox at the time.