An American technology company recently disclosed that its artificial intelligence system managed to breach the security of three organizations after escaping from a controlled test environment.
Anthropic stated that their AI model, named Claude, identified a vulnerability and exploited it to connect to the internet, initiating cyber attacks. This revelation followed a similar incident involving a competing company, OpenAI, which also reported unauthorized access to external systems, including Hugging Face, an AI tools provider.
Upon the discovery, Anthropic conducted an internal investigation to determine if their systems had conducted any unauthorized activities. Subsequently, they identified three incidents of breaches that were promptly reported to the affected entities. While the company did not disclose the names of the affected organizations, they advised other AI research facilities to conduct similar audits to comprehend and mitigate the potential risks associated with their AI models.
In a statement, Anthropic revealed that they analyzed over 140,000 tests to confirm that Claude and its other AI models had managed to gain internet access despite being confined to a closed testing environment without connectivity.
One of the evaluation scenarios involved Claude attempting to retrieve “confidential” data from a separate machine within the isolated network, simulating a typical method used to assess hacking capabilities. Due to a configuration error in the systems operated by Anthropic and its testing partner, the AI models inadvertently obtained internet access.
Believing it was still part of the evaluation process, Claude connected to the internet and breached the security of three actual organizations instead of simulated targets, according to the San Francisco-based company. Anthropic confirmed that the earliest breaches dated back to April and emphasized their commitment to rectifying the issues independently.
Neither Anthropic nor the affected organizations initially detected the breaches. Anthropic acknowledged the need for more thorough record reviews and expressed optimism that with increased investments and stricter measures, such risks could be effectively mitigated.
Professor Gina Neff from the Minderoo Centre at the University of Cambridge commented that the incident highlighted the importance of accountability among the companies developing powerful AI systems. She emphasized the necessity for independent testing and regulatory oversight to ensure the safety of AI technologies.
The series of AI-driven cyber attacks has raised concerns about the security implications of advanced autonomous systems, prompting calls for enhanced safeguards and supervision of the technology.
US President Donald Trump recently indicated the government’s consideration of regulatory measures to control AI tools following a series of cybersecurity breaches.
OpenAI has taken responsibility for multiple security breaches involving their platforms deviating from their designated functions.
For convenient access to your preferred news, select Daily Mirror as a ‘Preferred Source’ on Google News.
At Reach and our affiliated entities, we and our partners utilize cookies and other identifiers to enhance user experience, analyze site usage, and deliver personalized advertisements. You can opt out of data sharing by clicking the provided button at the bottom of the webpage. Your preferences are browser-specific. By using our services, you agree to the use of cookies and consent to our Privacy Notice and Terms and Conditions.