Safe & Secure AI Agent Practices
Anthropic Launches Enterprise Frontier Safeguards for Regulated Industries
Anthropic announced Enterprise Frontier Safeguards (EFS) on September 1, 2026, a security offering aimed squarely at a tension enterprise customers in regulated industries have been raising for a while: frontier models need retained activity data to catch misuse, but customers in finance, healthcare, and government often can't accept that data leaving their own infrastructure.
EFS resolves this by storing activity data in the customer's own cloud environment — AWS, Google Cloud, or Azure — rather than in Anthropic's systems, with the customer managing the encryption keys and access policies directly. Automated systems still analyze traffic patterns across sessions and accounts to catch serious threats like cyberattacks, credential theft, or offensive capability development, and alerts route straight to the customer's own security team rather than requiring an Anthropic human reviewer to see the underlying data. All of it is opt-in, with no changes to model behavior, pricing, or rate limits for customers who don't enable it.
Zero Data Retention is already available on Fable 5 and 5.1 ahead of EFS's general availability, and EFS itself begins a phased rollout this fall across Claude Code, Claude Enterprise, Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry. For teams building production agents in regulated environments, this is a direct answer to a real deployment blocker: it becomes possible to get frontier-model threat detection without shipping raw activity logs to a third party's servers.
Read more — Anthropic
Docker Details Six Concrete Benefits of Sandboxing AI Coding Agents
Docker published a detailed breakdown on September 8, 2026 of why sandbox environments matter specifically for AI coding agents, moving past the general "isolation is good" pitch into six concrete, actionable benefits. Each sandbox runs in its own microVM with a hardware-backed hypervisor boundary, so a compromised or misbehaving agent can't reach the host machine or other sandboxes running alongside it. Administrators can define network and filesystem policy — which domains, IP ranges, and host paths are reachable — enforced at runtime, blocking both unauthorized connections and data exfiltration attempts.
On credentials, the piece makes a specific architectural recommendation: credentials should stay in the host keychain and get injected into outbound requests at the sandbox boundary, rather than being passed to the agent directly through environment variables or mounted files, so the agent itself never has an opportunity to read, log, or leak them. Sandboxes are also designed to be ephemeral and disposable — fast to create and tear down, which enables running multiple agents in parallel without persistent state bleeding between runs — while still giving each agent a real Linux environment with package management and a full, isolated Docker daemon for building and running containers as part of its own workflow.
The piece's closing point is about consistency: running the same isolation model across every coding agent — Claude Code, Gemini CLI, Copilot CLI, and others — lets a platform team define governance once and have it apply uniformly, instead of maintaining separate security postures per tool. The concrete guidance: treat sandboxes as disposable infrastructure-as-code, never pass secrets directly, and centralize network/filesystem policy so it's enforced automatically rather than per-agent.
Read more — Docker Blog
Links & Sources
You May Also Enjoy
