Weekly signal

This week (Aug 17–25, 2026) privacy and security for agentic AI moved from research warnings into product and operational controls. Providers and defenders published new controls at three layers: model-provider privacy for enterprise (zero-data-retention + Private Safety Processing), national cyber guidance for agentic operational risk, and product releases that push enforcement to the network and endpoint. Key developments below show the emergent pattern: protect data at the model, isolate and observe agents in runtime, and stop risky actions before they execute.

What changed

1. OpenAI previewed Private Safety Processing (PSP) and reiterated Zero Data Retention (ZDR) for eligible frontier-model API customers, a design that aims to surface cross-interaction safety signals without exposing underlying customer prompts/responses to OpenAI personnel. OpenAI says PSP will be previewed with early customers and a technical whitepaper will follow. This is a practical industry attempt to reconcile enterprise data control with needed safety monitoring for long-lived agent tasks.

2. The UK’s National Cyber Security Centre (NCSC) published interim guidance on managing the cyber risk of agentic AI emphasizing threat modelling, sandboxing, observability/logging, human-in/on-the-loop controls, and an emergency "pull the plug" capability for agents. The guidance frames agentic systems as a new, high‑impact attack surface and gives immediate operational controls for defenders.

3. Vendors shipped agent-focused enforcement and supply-chain controls: Superagent released Context Guardrails and package/skill scanning to score untrusted web pages, files, skills, and packages before an agent consumes them; F5 announced enhancements to its AI Gateway for policy enforcement on model/agent traffic; Jamf and Norton shipped endpoint-level AI governance and agent-protection features for Mac/consumer devices. These show an industry move to gate agent inputs, tool installs, and actions at runtime.

What to do with it

• Treat this week as a governance inflection: prioritize threat modelling for agent use-cases, identify what data each agent can access, and require short-lived, task-scoped credentials at agent trust boundaries. (NCSC + product signals).
• Adopt multilayer controls: ZDR or equivalent for model access when possible, runtime sandboxes and observability, and pre-execution enforcement (agent action checkpoints / agent firewalls). Use providers’ ZDR/PSP details to negotiate auditability.
• Start red‑teaming agent supply chains and skills (scan packages, vet skill registries). Instrument agent telemetry into SIEM/SOAR and treat agent alerts like vulnerability reports.

Sources: numbered in the sources array below; see– for primary posts, product changelogs, and vendor releases.