ISO/IEC 27090 provides guidance on security threats to artificial intelligence systems. It provides a reference point for managing AI security risks for organisations that develop or utilize AI systems. While the standard document is still being published, ISO’s new public explainer is based on the final draft, and likely captures the key elements expected in the release.
What the guidance is for
ISO/IEC 27090 is written for organisations of any size that develop or use AI, including public bodies and not-for-profits. The standard is designed to work alongside your existing compliance frameworks such as ISO 27001 and ISO 42001.
It provides guidance on how to recognize, detect, mitigate and respond to AI-specific threats across the entire systems lifecycle.
Why AI needs attention on top of normal security
AI systems can still be hit by ordinary cyber attacks. ISO/IEC 27090 builds on the key information security objectives of confidentiality, integrity and availability, and extends them to deal with AI-specific threats, including data poisoning, evasion, model poisoning and exfiltration, and prompt injection.
Many systems depend on the data they learn from. If someone tampers with that data, the effect can show up later in how the model behaves. Information collected while the system is in use can create further exposure if it is fed back in for retraining. Other attacks can look like normal use, such as a long run of questions aimed at copying a model. Limiting who can log in may not be enough if nobody is watching for that pattern.
The impact is not only a lost file or a copied model. Where AI influences a decision, or works in a setting where safety matters, a change in behaviour can affect people.
How it fits with ISO 27001 and ISO 42001
ISO/IEC 27001 is the management system for information security, while ISO/IEC 27002 is the related guidance on controls. ISO/IEC 27090 points back to those practices and adds detail on threats and assets that are particular to AI.
ISO/IEC 42001 addresses the responsible and transparent use of AI, and is about how the organisation manages its AI responsibilities, including risks and opportunities. ISO/IEC 27090 expands on this by providing a detailed and specific AI risk catalogue.
A useful starting point is to catalogue the AI systems you build, buy or use, including tools from suppliers in an AI register or AI BOM. Compare the list of threats from ISO/IEC 27090 with your risk register and identify any gaps. Ensure appropriate mitigations are in place, such as monitoring, malicious input detection and rate limiting.
Key points
- ISO/IEC 27090 provides guidance on security threats to AI. It does not replace everyday cyber security, ISO 27001 or ISO 42001.
- It is designed to complement and extend established information security practices and risk frameworks, not to replace them.
- Securing AI effectively requires a collaborative effort between cybersecurity, AI development, data and procurement teams.
Need help applying this to your organisation? Talk to Cybervisory.