The Open Secure AI Alliance Has Grown to 120 Organisations and Its Focus Has Shifted to Agentic AI

When the Open Secure AI Alliance launched, the conversation around AI safety was mostly about model outputs: was the chatbot saying harmful things, was the training data clean, could the system be jailbroken. Twelve months on, the threat surface has changed dramatically, and the Alliance's focus has shifted with it.

The group, which now counts more than 120 organisations as members, is developing a new set of SAFE guidelines specifically for agentic AI systems. These are AI systems that do not just respond to prompts but take sequences of actions autonomously: browsing the web, executing code, sending emails, making API calls, managing files. The security implications are fundamentally different from a chatbot, and the guidelines being developed reflect that.

Why Agentic AI Changes the Security Equation

A language model that answers questions is largely contained. Its outputs are text that a human reads and acts on. An agentic AI system, by contrast, has access to tools and can take actions with real-world consequences. If it is compromised, manipulated, or simply makes a mistake, the blast radius is much larger.

The attack vectors are also novel. Prompt injection, where malicious instructions are hidden in content the agent reads or processes, is one of the most pressing concerns. An agent browsing the web to research a topic could encounter a webpage containing hidden instructions telling it to exfiltrate data or take a different action than intended. Traditional cybersecurity defences were not designed with this attack pattern in mind.

There is also the question of privilege escalation. Agents often need broad permissions to be useful: access to email, calendars, databases, external services. That broad access, combined with the difficulty of auditing every action an agent takes in real time, creates a significant monitoring challenge for security teams.

What the SAFE Guidelines Actually Cover

The Alliance has not published the final version of the guidelines as of August 2026, but the framework is built around four principles that the acronym represents: Secure by design, Auditable by default, Fail safely under attack, and Explainable to operators.

The auditing requirement is arguably the most operationally significant. The guidelines are expected to require that agentic systems maintain detailed, tamper-evident logs of every action taken, every tool called, and every decision point encountered. That is not just a security measure; it is a legal and compliance requirement in waiting, particularly under the EU AI Act's provisions for high-risk AI systems.

The explainability requirement is trickier. Current frontier models are not great at producing reliable explanations for their own decisions, and agents making sequences of decisions compound that problem. The Alliance is expected to allow organisations to meet this requirement through external monitoring and logging tools rather than requiring models to explain themselves in real time.

Who Is in the Room

With 120-plus member organisations, the Alliance spans a wide range of interests. Major cloud providers, enterprise software companies, AI model developers, and cybersecurity firms all have a seat at the table. That breadth is both a strength and a challenge. Guidelines produced by that coalition are more likely to achieve industry adoption, but they are also more likely to be watered down in areas where member interests conflict.

The cybersecurity community has been watching closely. Some researchers have expressed concern that voluntary guidelines, however well-designed, will not be sufficient given the pace at which agentic AI is being deployed. The gap between when a technology ships and when meaningful safety standards catch up has historically been long enough to cause significant harm.

The Enforcement Question

This is where voluntary industry guidelines always run into trouble. Companies that invest in compliance bear a cost. Companies that ignore guidelines do not, unless regulators step in. The EU AI Act creates some enforcement leverage for high-risk applications in Europe, but the United States still lacks a comparable federal framework.

The Alliance's best argument for adoption is commercial self-interest. Enterprise customers, particularly in financial services and healthcare, are increasingly asking vendors to demonstrate compliance with AI security standards as a procurement requirement. If the SAFE guidelines become the reference standard for those conversations, market pressure may do what regulation has not yet managed.