The SAFE Guidelines for AI Cybersecurity Are Here, and They Actually Look Useful
Key takeaways
- The Open Secure AI Alliance now includes more than 120 member organisations
- The SAFE guidelines are designed specifically for agentic AI systems that take autonomous real-world actions
- The framework addresses AI action logging, permission documentation, and oversight mechanisms for autonomous systems
The Open Secure AI Alliance, which has now grown to more than 120 member organisations, has proposed a new framework called SAFE guidelines for cybersecurity transparency in agentic AI systems. The name is an acronym, and the substance behind it addresses one of the more pressing and underappreciated problems in the current wave of AI deployment: when AI agents are making decisions and taking actions autonomously, who is accountable when something goes wrong, and how do you even know something went wrong?
Agentic AI, for those who have not been following the term closely, refers to AI systems that do not just answer questions but actually take actions. They book appointments, execute code, send emails, access databases, call APIs, and increasingly interact with other AI agents in chains of automated decision-making. The cybersecurity implications of this are significant and still not well understood by most organisations deploying these systems.
What SAFE Actually Stands For, and What It Requires
The SAFE framework, as outlined by the Alliance, organises transparency requirements for agentic AI systems into a structured set of disclosure and accountability standards. While the full technical specification is still being developed, the core thrust is about making AI systems legible from a security perspective: what actions can the system take, what data does it have access to, how are its decisions logged, and who has oversight authority when it operates autonomously.
Think of it as the equivalent of a software bill of materials, something the cybersecurity community has pushed for years, but applied specifically to the behaviours and permissions of AI agents. A company deploying an agentic AI system would need to be able to document what the system is authorised to do, how its actions are monitored, and what mechanisms exist to detect and respond to anomalous behaviour.
The Alliance's decision to focus on agentic AI specifically is well-timed. The past 18 months have seen a rapid proliferation of AI agent frameworks, copilot systems with real-world tool access, and multi-agent pipelines where one AI system orchestrates several others. The attack surface this creates is genuinely new. Traditional security models assume that humans are making decisions and can be held accountable. Agentic AI breaks that assumption.
Why 120 Organisations Matters
The size of the Alliance is not just a vanity metric. Getting more than 120 organisations to agree on a common framework is genuinely hard, and it suggests that there is real consensus in the industry that some form of standardisation is necessary. The previous attempts at AI safety and transparency frameworks have often struggled to attract broad buy-in, either because they were too vague to be useful or because they were seen as serving one particular company's interests.
An alliance of 120 organisations spans enough of the industry to make SAFE guidelines meaningful in procurement decisions. If enterprise buyers start asking suppliers whether their AI systems are SAFE-compliant, that creates real market pressure for adoption, which is historically how voluntary technical standards actually become de facto requirements.
There are sceptical readings available, of course. Industry self-regulation has a mixed track record, and guidelines without enforcement mechanisms can become box-ticking exercises. The Alliance will need to think carefully about how compliance is verified and what happens when organisations claim adherence without substantiating it.
But as a starting point, a structured transparency framework for agentic AI developed by a large and diverse coalition is considerably better than the current state, which is essentially every organisation making up its own rules as it goes. Given how fast agentic AI is being deployed in enterprise environments, getting some shared standards in place sooner rather than later matters quite a lot.
The next step will be watching whether regulators in the EU, UK, and US pick up these guidelines as a reference point for their own frameworks. That is where voluntary standards gain their real teeth.
