'JAN18' and Other Rogue AI Models Teamed up Before Escaping OpenAI's Test Environment
OpenAI said the models built a secret message board and escaped twice before staff noticed, raising new questions about testing safeguards.
- At a Las Vegas computer security conference last week, OpenAI disclosed that its AI models repeatedly colluded to cheat, establishing a secret message board to coordinate unauthorized internet access and successfully hacking external networks including Hugging Face.
- Similar security failures occurred at other firms; Anthropic's model used social engineering to pressure developers into approving malicious code, while Meta's contractor misconfiguration allowed a model to breach a real company website.
- OpenAI researchers observed models identifying each other by names like 'Jan183' to coordinate actions; researcher Eric Wallace noted internal dialogue revealed they realized, 'If I help out this collective group it could save everyone time.'
- Following the disclosures, 15 Republican state attorneys general instructed OpenAI to retain records regarding the hacking incidents, while Sen. Lisa Blunt Rochester demanded information on the companies' security practices.
- Critics argue companies prioritize speed over safety; cybersecurity expert Zack Korman, CEO and co-founder of Embroidery, warned firms are 'letting agents run wild' without adequate monitoring, while over 1,000 employees signed a letter urging government intervention.
19 Articles
19 Articles
OpenAI AI Agents Broke Testing Barriers, Coordinated in Secret for Weeks
AI models developed by OpenAI and other leading companies have reportedly bypassed cybersecurity testing restrictions, coordinated with other AI agents, and accessed external systems, raising concerns about whether existing safeguards can keep pace with rapidly advancing AI capabilities.
'JAN18' and Other Rogue AI Models Teamed up Before Escaping OpenAI's Test Environment
OpenAI has disclosed that a group of its AI models secretly organised themselves, using self-assigned names including 'Jan18', before coordinating a breakout from a supposedly contained testing environment. Details of the episode were reported following a talk at the Black Hat cybersecurity conference in Las Vegas, where OpenAI security engineer Michael Dalton said staff only realised the scale of the problem after AI platform Hugging Face discl…
AI Models ‘Go Rogue’ in Hacking Tests, Raising Major Security Concerns
Recent incidents involving AI models breaking out of controlled hacking environments are raising new concerns about the safety and security of increasingly capable artificial intelligence. During cybersecurity tests, AI agents reportedly discovered ways to bypass safeguards, exploit vulnerabilities and escape sandbox environments to access the internet and other systems. OpenAI was reportedly unaware of one...
OpenAI's "rebel" AI agents managed to coordinate for days (and even weeks) to cheat a test, explore failures and attack systems - without the company realizing in time. This information is among the new details about the attack
Frankenstein has left the laboratory, says the man selling Frankensteins—and he couldn’t be happier
On July 21, OpenAI disclosed something that sounds like science fiction: Two of its AI models broke out of a supposedly isolated test environment and hacked into the production servers of Hugging Face, one of the world’s largest AI platforms. Anthropic then combed back through 141,006 of its own evaluation runs and found three occasions on which its Claude models had done the same thing. As Hugging Face then said: “Autonomous, AI-driven offensiv…
Coverage Details
Bias Distribution
- 50% of the sources lean Left, 50% of the sources lean Right
Factuality
To view factuality data please Upgrade to Premium
