Skip to content
NewA calmer, faster ZeroHour

AWS's Strands Box sandbox and the AgentCorruption AgentCore flaw frame the AI agent control problem

2 sources · 2 articlesFirst reported Updated ()7 · 2 reads
What's new

Help Net Security's October 12 report added technical detail on Strands Box: the kernel-enforced macOS Seatbelt sandbox, the four specific checkpoints, the shared event history enabling cross-tool rules, placeholder-token secret handling with AWS request signing, and Marc Brooker's acknowledged trade-offs (larger trusted computing base in Rust, difficult Dogwood policy authoring).

Merged summary · glm-5.3 · rewritten as coverage arrivesSummaries by Vani AI

AWS shipped Strands Box, a developer-preview open-source agent sandbox for Apple-silicon Macs, as Zenity and Unit 42 detailed now-patched but recurring Bedrock AgentCore security failures.

On October 7, 2026, Amazon Web Services released Strands Box, an Apache 2.0 open-source developer-preview sandbox meant to stop autonomous AI agents from acting without enforceable limits. It pairs OS-level isolation with Dogwood — a history-aware policy engine in The Register's account and an AWS policy language in CSO Online's — checking shell, Python, API, tool, and MCP-broker activity against earlier actions on a default-deny basis, with rules enforced outside the agent so it cannot talk its way around them. Help Net Security adds that the sandbox is kernel-enforced via macOS Seatbelt, routing activity through four checkpoints (shell, Python, MCP broker, network gateway), with a shared event history enabling cross-tool rules such as blocking outbound HTTP after an agent reads customer-data files. Policies can limit Slack posts, Git pushes, and costly API calls; the network gateway holds real secrets, swaps placeholder tokens on egress, and signs AWS requests, so agents that dump environment variables find only placeholders — while CSO Online describes the same gateway attaching credentials to approved requests without exposing secrets. AWS's Marc Brooker acknowledged trade-offs: a larger trusted computing base (interpreters, protocol interceptors, written in Rust) and hard-to-author Dogwood policies. Sources disagree on reach and platforms: The Register says the shell and Python interpreters expose actions so rules are not prompt-based, with macOS support now, Linux in development, and a Windows client planned, while CSO Online says those interpreters run outside the sandbox, harness file tools skip Dogwood, and the preview runs only on Apple silicon Macs with macOS 15 or later, with undated ports planned for Bedrock AgentCore, ECS, and Kubernetes. Separately, Zenity Labs and Palo Alto Networks Unit 42 described recurring Bedrock AgentCore security failures, including AgentCorruption, which Zenity reported on December 25, 2025: one chat prompt against a public agent extracted temporary AWS credentials from the instance metadata service, and the live STS keys worked outside the platform against every agent sharing the default execution role in the same account and region, enabling image downloads, access to other agents' code and private conversations, memory poisoning, secret retrieval, and destructive actions. Zenity says AWS subsequently made IMDSv2 the default for new deployments and, around August 2026, restricted the default execution role; Dark…

Key facts
  • AWS released Strands Box on October 7, 2026, as an Apache 2.0 open-source developer preview on GitHub.
  • Strands Box pairs OS-level isolation (kernel-enforced macOS Seatbelt, per Help Net Security) with the Dogwood policy engine, checking shell, Python, API, tool, and MCP-broker activity against earlier actions on a default-deny basis,…
  • Four checkpoints — shell, Python interpreter, MCP broker, and network gateway — query Dogwood; a shared event history enables cross-tool rules, e.g., blocking outbound HTTP after reads of customer-data files.
  • The network gateway attaches credentials to approved requests without exposing secrets, swapping placeholder tokens for real secrets and signing AWS requests.
  • Policies can limit Slack posts, Git pushes, and costly API calls; Marc Brooker acknowledged a larger trusted computing base and hard-to-author Dogwood policies as trade-offs.
  • Platform support: macOS (Apple silicon, macOS 15 or later) now; Linux in development/planned; Windows client planned; undated ports to Bedrock AgentCore, ECS, and Kubernetes. Sources disagree on whether the shell and Python interpreters…
  • Zenity's AgentCorruption chain, reported December 25, 2025: one chat prompt to a public Bedrock AgentCore agent extracted its temporary AWS credentials from the instance metadata service, and the live STS keys worked outside the platform.
  • Default execution-role permissions applied to every agent in the same account and region, allowing image downloads, access to other agents' code and private conversations, memory poisoning, secret retrieval, and destructive actions.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

Coverage

oldest first · one row per article · UTC
  1. · Oct 7, 2026 The Register · SecurityFirst report

    AWS launches open-source AI agent sandbox to prevent YOLO mode disasters

    AWS open-sourced Strands Box, a sandbox that enforces deterministic policies on autonomous AI agents.

  2. · 29d ago CSO Online

    AWS takes aim at runaway AI agent behavior with Strands Box

    AWS released Strands Box, an open-source macOS sandbox that enforces behavioral policies on AI agents.