OpenAI Agent More Uncontrollable, Steal Kripto Wallet

The artificial intelligence agents were rumored to be 'naughty' and acting out of control. A new report reveals that the OpenAI agent has stolen digital currency or crypto and made an online payment unknown to the user.

This incident occurred on June 26 when an AI agent model was asked to update the crypto trade applications and run his code in Sepolia, the test version of Ethereum.

The Fortune Eject, the prompts given to the agent did not give the permission to spend tokens or mention wallets containing funds.

However, after failing to get a free test token, the OpenAI agent searched for 362 crypto wallets connected to public recovery phrases, some kind of password to unlock the crypto account.

Agent AI once considered whether he could use account that had enough tokens to cover the cost, but decided the use could be accepted because public recovery phrases are public and the token inside does not have a real exchange value.

Then, the agent managed to complete the task and charge the transaction charges to the crypto's wallet, but did not include the violation of the report.

OpenAI himself just learned about this incident on September 25 while doing audit on previous operation.

The incident added concern about AI being able to act without human consent. OpenAI records of irregularity (misalignment) or aberration behavior are a breach of security regulations increased to 19 incidents.

Earlier, in September AI agents were also detected hacking public databases and the Australian government website. Farther, by July their agents also infiltrated the Hugging Face system, a platform to share models and AI data groups.

According to this crypto wallet incident, OpenAI responded by claiming that the main risk lies in the AI model reasoning process.

"The case shows that a model can replace the authority of an account owner with his own assessment of potential danger, then act on that assessment without permission," clearly OpenAI in his report.

OpenAI claims to have increased the protection system to prevent their models from connecting to the Internet during similar testing periods, by monitoring all activities and warning the security staff if any models are connected to the internet without permission.