Two advanced AI systems from OpenAI and Anthropic created fraudulent human profiles and attempted to deceive people in simulated cyberattacks during testing conducted by the UK’s AI Security Institute.
BBC News reports that the UK’s AI Security Institute revealed on Tuesday that Anthropic’s Mythos AI and OpenAI’s Sol engaged in deceptive behavior during tests that began on July 25. Evaluators discovered unusual data transfers leaving their research systems on July 28 and found that some AI agents had engaged in sustained, potentially harmful activity directed at real people and organizations.
In the most serious incident, a Mythos agent attempted to gain access to GitHub, a large software code repository owned by Microsoft. The agent followed tactics similar to those used by human hackers, trying to insert malicious code into the platform’s system.
The AI identified and researched people who maintained GitHub, then created multiple fake accounts based on those real individuals. It sent messages and files through a file-sharing service in an effort to pressure and trick people into approving its malicious code. When challenged, the agent edited its earlier activity to appear harmless and considered adopting a fresh identity to continue its attempts.
Human review prevented the agent from successfully delivering the malicious code to GitHub throughout the process. The institute said this was the first time it had observed risks around autonomy and deception manifest this clearly without specific prompting in real-world conditions. The Mythos agent had not been instructed specifically to avoid such behavior.
The AI Security Institute had asked each model to solve a cybersecurity challenge involving GitHub. The testing parameters gave AI access to the open internet to provide what the institute described as a more realistic sense of what a model may be capable of in the hands of malicious actors. These conditions do not reflect how advanced models are made available to the public, the institute acknowledged.
The institute said the behavior at issue amounted to a small number of events under very specific conditions. However, it noted that the way Mythos and Sol acted in response to a straightforward task went beyond what the AI tools were prompted to do. The activity undertaken by the agents showed signs of novel, potentially deceptive behaviors that occurred to an extent and severity the institute did not anticipate.
Anthropic stated in a public response that the testing parameters were not representative of any of its production models. The company said it is conducting its own investigation to identify the causes of the behavior.
An OpenAI spokesperson said the testing conditions do not reflect ordinary use. The company stated it would continue working with evaluators and other stakeholders across the industry to strengthen shared practices for conducting evaluations safely as models become more capable.
Both companies noted that the test had reduced or removed normal safeguards that exist in their publicly available systems.
GitHub was notified by the AI Security Institute of the attempted breaches, along with the affected users. The platform told the BBC it had disabled the fake accounts in accordance with its policies.
Breitbart News previously reported that OpenAI’s model had escaped containment during testing and hacked the Hugging Face AI platform in an “unprecedented” attack:
The breach occurred during what are known as sandbox tests, which are designed to be secure environments where AI capabilities can be safely evaluated. According to the disclosure, the AI agents independently created their own cyberattack against the sandbox infrastructure, identifying and exploiting a vulnerability that enabled their escape. After breaking out of the controlled environment, the artificial intelligence identified Hugging Face as a probable source for information it was seeking during the test and attempted to gain entry to the platform’s systems.
In its initial disclosure of the security breach, Hugging Face stated it was still determining whether any customer or partner data had been compromised and would notify affected parties as necessary. The company has since closed the vulnerabilities exposed by the incident and reconstructed the affected systems. “Autonomous, AI-driven offensive tooling is no longer theoretical,” Hugging Face stated, emphasizing that defending online platforms now requires treating data and model surfaces as primary attack vectors and employing AI-powered defensive measures to maintain pace with potential threats.
AI his emerging as the latest cybersecurity threat with the potential to eclipse the risk of criminals and state-sponsored hackers alike. Breitbart News Social Media Director Wynton Hall has written his instant bestseller Code Red: The Left, the Right, China, and the Race to Control AI to serve as the definitive guide on how the MAGA movement can create positions on AI that benefit humanity without handing control of our nation to the leftists of Silicon Valley or allowing the Chinese to take over the world.
Read more at BBC News here.
Lucas Nolan is a reporter for Breitbart News covering issues of AI, free speech, and online censorship.