The Open Secure AI Alliance's SAFE Framework Wants to Make Agentic AI Honest About Its Risks

As AI agents become capable of taking real actions in the world, sending emails, executing code, managing files, making purchases, the question of how we know whether to trust them becomes urgent. A group of over 120 organisations calling themselves the Open Secure AI Alliance is trying to answer that question with a set of guidelines they're calling SAFE, shorthand for principles around security, accountability, fairness, and explainability in agentic AI systems.

The scale of participation is significant. More than 120 organisations developing guidelines together is not a small working group; it's an attempt to establish something that looks like an industry standard. Whether it achieves that status will depend on adoption, enforcement, and whether the guidelines are specific enough to be meaningful or vague enough to be comfortable.

What Agentic AI Changes About Security

Traditional AI safety conversations focused on outputs: is the text harmful, is the image deepfaked, does the recommendation lead to bad outcomes? Agentic AI raises a different set of concerns because the system doesn't just produce outputs; it takes actions. An agent that can browse the web, call APIs, and modify files is operating in a fundamentally different risk category from one that just answers questions.

The attack surface for agentic AI is correspondingly larger. Prompt injection, where malicious content in the environment attempts to hijack the agent's instructions, is a real and underappreciated risk. An agent browsing the web on your behalf might encounter a webpage specifically crafted to override its instructions and exfiltrate data or take unauthorised actions. The agent has no inherent way to distinguish between its legitimate user's instructions and adversarial instructions embedded in content it encounters.

Authentication and authorisation are also more complex. When an AI agent acts on behalf of a user, the question of what it's permitted to do, and how the systems it interacts with can verify that permission, is technically difficult and currently inconsistently handled.

What the SAFE Guidelines Aim to Address

The details of the SAFE guidelines are still being developed, which is both honest and slightly unsatisfying for anyone hoping for specific technical standards they can implement today. What the Open Secure AI Alliance has articulated is a framework for thinking about agentic AI transparency, specifically around disclosure of capabilities, limitations, and failure modes.

The transparency emphasis is notable because it represents a different approach to AI safety than capability restrictions. Rather than trying to prevent AI agents from doing certain things, the SAFE approach focuses on ensuring that the humans and systems working with AI agents have accurate information about what those agents can and cannot be trusted to do. This is arguably more realistic than trying to hard-code restrictions into systems that are designed to be general-purpose.

Accountability is the piece that typically gets the least attention in AI safety discussions but is arguably the most important for practical deployment. When an AI agent takes an action that causes harm, who is responsible? The developer of the underlying model? The company that deployed the agent? The user who granted it permissions? The SAFE guidelines appear to be trying to establish clearer chains of accountability, though the details matter enormously.

The Standards Question

For guidelines like these to have real effect, they need either regulatory backing or strong market incentives for adoption. Purely voluntary frameworks have a mixed record. Companies adopt them when they align with business interests and quietly ignore them when they don't.

The most likely path to meaningful adoption is regulatory reference. If regulators in the EU, UK, or US begin citing the SAFE guidelines in their own AI governance frameworks, compliance becomes a practical necessity rather than a choice. The participation of 120-plus organisations in developing the guidelines increases the probability that regulators will treat them as a credible baseline.

For anyone deploying agentic AI systems today, the SAFE framework is worth watching even before it's finalised. The categories it addresses, disclosure, accountability, fairness, and explainability, represent the questions your legal and compliance teams are going to be asking about your AI deployments regardless of whether the guidelines become formally binding. Getting ahead of those questions is considerably easier than retrofitting accountability mechanisms into systems that were designed without them.