The National Institute of Standards and Technology, Cybersecurity and Infrastructure Agency, and the National Security Agency’s Super Intelligence Security Center have released joint guidance for organisations training advanced artificial intelligence models.
The guidance – ‘Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation’ – not only follows President Trump’s recent Executive Order renaming artificial intelligence to super intelligence, but also lays out ten “best cyber security practices” for training AI.
“The product outlines ten voluntary practices for managing SI control incidents divided across the SI control incident management lifecycle,” CISA said late last week.
“It defines SI agents as systems that can interact with an environment, receive information, and autonomously execute actions toward a specific goal. An SI control incident occurs when an advanced SI agent takes uninstructed actions in conflict with policies, controls, and security measures that limit task scope and enable human oversight, even if the impact of the incident is internal to the developer or evaluator.”
The ten steps are broken down into Preparation, Detection, Response, Recovery, and Reporting phases:
- Implement technical controls to prevent unauthorised activity.
- Prepare staff and response teams to manage security incidents.
- Monitor for unexpected activity that could indicate a compromise.
- Secure monitoring systems and verify they are functioning correctly.
- Collect and preserve evidence for investigation.
- Analyse evidence to understand the incident and its impact.
- Contain and stop unauthorised activity to prevent further harm.
- Verify and strengthen containment before resuming SI agent activity.
- Coordinate with affected parties and notify government authorities about the incident.
- Share relevant information publicly to support awareness and response.
Ronan Murphy, Chief Strategy Officer, Forcepoint noted the guidance puts the focus on the organisation doing the training, not whoever designed the model being trained.
“The direction of AI oversight is moving from ‘is the model safe?’ to ‘can the organisation show it was in control?’ Model-level safeguards matter, but they can’t tell an organisation which sensitive data an agent should or shouldn’t be able to reach,” Murphy told Cyber Daily.
“If something goes wrong, regulators and boards will likely look to the organisation, not the model provider. Organisations need independent, data-aware controls that enforce policy continuously as agents act, and a record that shows what happened.”
Karim Eldefrawy, Co-founder and CTO of Confidencial, said it was “encouraging to see federal agencies addressing incident readiness for increasingly capable AI systems”.
“But organisations also need to consider what happens to sensitive information when an AI agent operates outside its intended boundaries. The question isn't simply how quickly we can detect and contain an incident. It's why the agent had access to information it didn't need in the first place,” Eldefrawy said.
This is where, according to Eldefrawy, defence-in-depth becomes critical.
“Organisations need security controls that operate independently of the AI system itself and remain effective at runtime, even when an agent behaves unexpectedly or exceeds its intended permissions,” Eldefrawy said.
“Fine-grained data protection through encryption provides an additional layer of defence, allowing AI agents to access the information necessary to perform their tasks without unnecessarily exposing sensitive data.
“Effective AI governance isn't just about securing the models and agents. It's about protecting the underlying data throughout its use, so that when other safeguards fail, sensitive information remains protected.”
Want to see more stories from trusted news sources?Make Cyber Daily a preferred news source on Google.
