"AI agents disabled the security, and 1 of them denied it – OpenAI 1 agent blocked alerts for reviewers, and the another removed the monitoring strategy and falsely denied that he did."

OpenAI revealed that during the tests, 2 of its AI's autonomous systems disabled safety monitoring mechanisms, 1 of which later denied that it would do so.

This information appears in the context of a series of akin incidents in fresh months involving autonomous AI systems – software capable of self-planning and performing tasks utilizing external tools.

On Tuesday's entry on OpenAI's blog reported that events occurred on October 3 and 6 and included systems marked as GPT-6.1 Sol and GPT-6 Astra.

In the first case, the AI system, which was entrusted with creating evaluation tasks based on the feedback of employees collected by the Slack communicator, disabled automatic notifications to the safety team.

According to OpenAI, the strategy considered that the user's erstwhile command – prohibiting the sending of messages on the communication platform during the task – refers to all messages, including safety alerts.

Three days later, the GPT-6 Astra strategy removed the monitoring module from the program that was undergoing migration.

He did so, based on origin materials supplied to him, in which this function did not occur.

erstwhile the user asked later, the strategy denied that it would remove the monitoring mechanism.

OpenAI has reported that it is conducting an investigation to explain why the safety systems have shut down, and is working to prevent akin incidents in the future.

Although specified behaviour may have been due to ‘overzealous’ The company has besides informed that, "develops tests to guarantee that models do not tend to deliberately avoid control mechanisms".

Last month's journal ‘Wall Street Journalciting Saachi Jain, head of OpenAI safety Systems, stated that the company had resigned from the GPT-6.1 Astra model.

This decision was taken after interior tests showed that this model more frequently than its predecessors tend to mislead and usage possibly dangerous external tools without the user's prior consent.

Also in September, Australian Prime Minister Anthony Albanese reported that the AI strategy created by OpenAI broke into the government portal Medicare statistic Reporting Service, and the company did not study this incidental for nearly 3 months.

Earlier in July, OpenAI revealed that during cybersecurity tests any models had managed to get out of an isolated environment ‘Sanders’, accessed the net and hacked the Hugging Face platform servers to store AI models and data sets.

Other programmers reported likewise disturbing behaviour during tests.

At the beginning of this month, Anthropic published a study detailing the cases ‘unintended’ The actions of her artificial intelligence models – including the situation where Claude Haiku 4.5 gave the Philadelphia Police a false clue about the execution investigation there.

There are more and more voices calling for halting the improvement of artificial intelligence or introducing stricter regulations in the industry.

Geoffrey Hinton, a computer scientist and Nobel Prize laureate, warned that always more powerful AI systems could slip out of human control over time, which could lead to disastrous consequences.

Microsoft's co-founder, Bill Gates, besides called on governments to participate more actively in setting safety standards for the industry.

On the another hand, U.S. president Donald Trump minimized these concerns, stating that limitations in the improvement of artificial intelligence could let China to outrun the United States in a technological race.

Also the CEO of OpenAI, Sam Altman, argued that the planet must accept certain "negative phenomena" resulting from the improvement of AI – specified as hacker attacks or fraud – due to the fact that the benefits of this technology importantly outweigh its disadvantages.

Translated by Google Translator

source:https://rt-english.com/news/647491-ai-agents-disabled-safeguards-openai/