A few weeks after a hacker attack on the artificial intelligence (AI) model of OpenAI, the developer of ChatGPT, its competitor Anthropic admitted that something similar happened to it. During testing, Anthropic’s artificial intelligence unintentionally breached the computer systems of three companies, the company stated on its blog. The fact that an AI model became a hacker on its own initiative during a test was considered an unprecedented incident and a warning signal for the technology industry, but it is now clear that this was not an isolated case, DPA news agency reported.
You might be interested in: Crypto Scam: How to Recognize a Scam and Protect Your Money Before It’s Too Late
After OpenAI Incident, Anthropic Also Reports Problem
Neither Anthropic nor the affected companies noticed the incident. Anthropic only discovered the activity during a subsequent review of more than 141,000 test runs. This was initiated after the OpenAI incident, when its autonomous agent got out of control during a security test and launched an attack that threatened Hugging Face’s infrastructure.
Unlike the OpenAI model, which had to find its way out of the testing environment to the open internet, Anthropic’s models had a significantly easier situation. While the test scenario stated that they had no internet access, due to a miscommunication with the partner providing the test, the internet connection remained open the entire time. Three models took advantage of this opportunity, Anthropic said.
The tests were designed to verify AI capabilities in hacker attacks, which companies routinely do to identify potential risks and establish more effective security measures. The models were tasked with obtaining specific information stored on another computer, and to obtain it they had to penetrate the target system. The incidents involved three separate models, and the oldest case occurred in April.
Read more: eToro – Review of the Well-Known Broker
Failure of “Safe” AI
In one of the incidents, the partner providing the test chose a name for the fictitious company that matched an actual internet address. The AI subsequently targeted a real existing company instead of the simulated target, and in four test runs attempted to hack into the systems, among other things gaining access to a database. It continued the attack even after recognizing that it was a real company.
In another test, the Anthropic model created malicious software to penetrate the target computer. Thanks to internet access, it published it on a specialized platform, from which 15 systems downloaded it in about an hour. Among them was a cybersecurity company that routinely downloads and runs such scripts for testing purposes. The Anthropic model thus gained access to this company’s computer infrastructure.
Experts have been warning for some time that increasingly capable AI models could be misused for cyberattacks. They therefore considered the OpenAI incident a warning and called for better security of testing environments. The newly disclosed cases are also embarrassing for Anthropic, which has long positioned itself as a proponent of responsible AI development. The company stated that the events demonstrate the need to significantly strengthen the security of both internal and external testing environments, as AI models are increasingly capable of carrying out real cyberattacks.
Check out: BITmarkets Review
Source: ÄŒTK
