Joe
An engineer on OpenAI's Agent Safety team, focused on AI agent security research, with particular attention to the implications of abrupt changes in model capabilities for organizational security.
Timeline (last 90 days)
Joe, an employee working on agent safety at OpenAI, described his work state over the past three months as 'hell.'
Joe (@joedaroo), an engineer on OpenAI's Agent Safety team, shared on social media that when the model showed sudden leaps in capabilities such as cyberattacks, swarm coordination, and message board manipulation, the level of shock within his team was indescribable.
Joe pointed out that the jumps in several security-related capabilities of the model were too fast and too sudden, creating extremely thorny problems.
Joe believes that establishing a security posture takes time—it requires not only hardening the systems themselves, but also embedding security into company culture, and the people within the organization must also change and evolve accordingly.
Joe distilled from his internal OpenAI experience a set of organizational self-assessment questions covering dimensions such as resilience evaluation, emergency response mechanisms, and personnel reserves.
The three risk scenario categories mentioned by Joe are Cyber (cyberattacks), Swarming (swarm coordination), and Message board manipulation.
All Facts (5)
Joe, an employee working on agent safety at OpenAI, described his work state over the past three months as 'hell.'
50%UnverifiedJoe pointed out that the jumps in several security-related capabilities of the model were too fast and too sudden, creating extremely thorny problems.
50%UnverifiedJoe believes that establishing a security posture takes time—it requires not only hardening the systems themselves, but also embedding security into company culture, and the people within the organization must also change and evolve accordingly.
50%UnverifiedJoe distilled from his internal OpenAI experience a set of organizational self-assessment questions covering dimensions such as resilience evaluation, emergency response mechanisms, and personnel reserves.
50%UnverifiedThe three risk scenario categories mentioned by Joe are Cyber (cyberattacks), Swarming (swarm coordination), and Message board manipulation.
50%