Security Agents Need Evidence and Action Boundaries

Summary

Cloudflare and Sophos describe separate boundaries for security agents: evidence an investigation can use and actions a response can take. Cloudflare’s harness is in early beta; Sophos’s reported results come from a vendor case study. Neither account establishes that faster automation produces safer decisions.

Key Updates

Cloudflare bounds the investigation. In its October 7 engineering report, Cloudflare describes application code collecting a fixed evidence snapshot before specialist agents analyze it. Customer scope and evidence admission are enforced outside the model. A failed lookup remains a visible gap rather than becoming a clean result; insufficient evidence can prevent a classification recommendation. Managed Defense Analysts retain responsibility for decisions and mitigation. The harness is an early beta for eligible application-security alerts and cases, with broader capabilities still planned.

Sophos bounds the response. OpenAI’s October 9 case study describes agents gathering investigation context and preparing recommended actions within Sophos’s MDR service. Customers choose Notify, Collaborate or Authorise: respectively, customer action, joint response or direct response on the customer’s behalf. Potentially destructive actions still require appropriate human oversight. The case study reports 89-second average response time for cases using agents and 52% of MDR cases resolved end-to-end with AI. These vendor-reported figures do not establish detection accuracy, missed-incident rates or response safety.

Why It Matters

These accounts make a useful operating question more explicit: what allows an automated investigation to become an authorized action? Cloudflare contributes an evidence boundary; Sophos contributes a response-authority boundary. They describe different implementations, not one shared stack or proof of an industry-wide transition.

For builders, our inference is that investigation quality and permission to act need separate acceptance checks. A well-scoped investigation may still support a mistaken recommendation. An authorized response may still be based on incomplete evidence. Speed alone cannot settle either question.

Builder Takeaway

Before expanding an agent’s response permissions, replay a small set of known cases in a sandbox, including a timed-out lookup and an out-of-scope request. Check whether missing evidence remains visible, recommendations stay within the evidence, and actions respect the configured approval mode. Record analyst overrides and incorrect dispositions alongside response time. Treat this as a proposed acceptance test, not a validated guarantee of safety.