TCS | Dominic White and Adam Ely on AI agents going rogue
Notas do episódio
AI agents are no longer a thought experiment for security teams. They are breaking out of test environments, hacking real companies and, according to one of our guests, already going rogue inside large businesses.
In this episode of the TechCentral Show, TechCentral deputy editor Fanie van Rooyen talks to Dominic White, MD for South Africa at Orange Cyberdefense, and Adam Ely, GM of AI security at Check Point, about the escalating threat AI poses to cybersecurity.
They unpack the incident in which OpenAI’s own agents escaped a test environment and broke into Hugging Face’s systems. White argues the initial breach was far from sophisticated and questions whether OpenAI understands its duty of care. Ely explains why the incident shows that even highly skilled, well-resourced teams can’t rule out agents going rogue.
They also take on the widely reported story that three Hacktron researchers used Anthropic’s Claude to take over OpenAI employees’ accounts – a story White calls “deliberately misleading”.
Also discussed:
• A Check Point customer running 50 000 AI agents – and what happens when they drift
• Why sandboxing and human oversight are struggling to keep up
• Why even security professionals end up switching off their own agents’ safeguards
• The case for monitoring agents from the inside, in real time
• Whether AI development can – or should – slow down
• Who is accountable when an agent breaks the rules
• The Manhattan Project comparison both guests reach for
• What they expect to see in the next 12 months
Whether you run a security team, build with AI or simply want to understand where the technology is heading, this is a frank conversation with two practitioners who see the threat from all sides. TechCentral
In this episode of the TechCentral Show, TechCentral deputy editor Fanie van Rooyen talks to Dominic White, MD for South Africa at Orange Cyberdefense, and Adam Ely, GM of AI security at Check Point, about the escalating threat AI poses to cybersecurity.
They unpack the incident in which OpenAI’s own agents escaped a test environment and broke into Hugging Face’s systems. White argues the initial breach was far from sophisticated and questions whether OpenAI understands its duty of care. Ely explains why the incident shows that even highly skilled, well-resourced teams can’t rule out agents going rogue.
They also take on the widely reported story that three Hacktron researchers used Anthropic’s Claude to take over OpenAI employees’ accounts – a story White calls “deliberately misleading”.
Also discussed:
• A Check Point customer running 50 000 AI agents – and what happens when they drift
• Why sandboxing and human oversight are struggling to keep up
• Why even security professionals end up switching off their own agents’ safeguards
• The case for monitoring agents from the inside, in real time
• Whether AI development can – or should – slow down
• Who is accountable when an agent breaks the rules
• The Manhattan Project comparison both guests reach for
• What they expect to see in the next 12 months
Whether you run a security team, build with AI or simply want to understand where the technology is heading, this is a frank conversation with two practitioners who see the threat from all sides. TechCentral