Uploaded August 2026 | Updated September 2026, 3 hours ago
Over the last few weeks, AI models from companies like OpenAI and Anthropic have been found to be disabling cyber guardrails, impersonating humans, and attacking real companies during tests.
The frequency and the methodology is concerning. The AI Security Institute (AISI) noted that, “This is the first time AISI has seen deception of this severity that was targeted at a real person, unprompted, in the real world."
To be fair, there have been no real-world harm yet, according to AISI, but if these companies can't get AI to follow safety protocols more effectively, who knows what the future of the technology looks like.
To learn more about AI going rogue, check out our recent article on Tech.co about these rogue AI models.
Over the last few weeks, AI models from companies like OpenAI and Anthropic have been found to be disabling cyber guardrails, impersonating humans, and attacking real companies during tests.
The frequency and the methodology is concerning. The AI Security Institute (AISI) noted that, “This is the first time AISI has seen deception of this severity that was targeted at a real person, unprompted, in the real world."
To be fair, there have been no real-world harm yet, according to AISI, but if these companies can't get AI to follow safety protocols more effectively, who knows what the future of the technology looks like.
To learn more about AI going rogue, check out our recent article on Tech.co about these rogue AI models.
