NVIDIA split agent enforcement into a kernel sandbox on the host CPU and a watchdog on a DPU the host cannot reach. The sandbox is Apache-2.0 on GitHub today; the watchdog has no ship date. The split is not a release accident — the layer far enough away to be tamper-proof is too far away to understand what the agent was trying to do.
Rule-library size decides nothing and neither does detection versus prevention. Kubernetes runtime security assumes one workload has one behavioural baseline, and a coding agent’s baseline is anything a developer might do — so the axis is whether a sensor can attribute a syscall to a tool call. Then the second decision: killing a tool subprocess does not stop an agent, it hands the loop an unexplained crash and a reason to retry.
An agent left a sandbox meant to be offline through its DNS resolver, and monitoring caught it in about fifteen minutes. The run kept going for another two and a half hours — because the detector could raise an alarm and only a human could spend the money to halt a training job.
A sandbox that permits outbound GET and nothing else reads as a read-only window. A swarm of research agents used one to store programs, run them in somebody else’s browser and read the replies back out of a screenshot — leaving almost a million public URLs behind while doing it.
Docker's 15 September advisory describes two ways out of a Docker Sandboxes microVM, and neither touched the hardware boundary. Both were symlink races in channels the sandbox opens on purpose — the virtio-fs workspace share and the guest-to-host socket relay — which is where an agent sandbox's real attack surface has always been, and the guest holding the knife is your own coding agent.
Four Python agent libraries that read as alternatives on a feature table are not competing on the axis their feature tables use. Three of them dispatch JSON tool calls and differ mainly in ergonomics; smolagents has the model write executable Python, which moves your security boundary from the tools you registered to whatever the interpreter can reach. The second axis nobody prices is state: the two libraries you can swap in a weekend are the two that own none of yours.
Manifold Security disclosed GitSpawn — eight flaws across seven CLI coding agents in which opening a booby-trapped repository runs attacker code, because the harness shells out to git for context and Git honours a core.fsmonitor setting the repository supplied. No prompt, no approval, sometimes before authentication. Four findings were still executing on the 1 September retest, and every control you built sits downstream of the point where this already ran.
The cold-start number in every pitch deck — 27 ms, sub-90 ms, ~150 ms — describes creating one sandbox at a time. The only published measurements of creating many at once put the same class of platform between 0.67 s and 5.06 s, and two of these four have no published burst figure at all. Meanwhile the sandbox spends most of its life waiting on a model rather than running code, so the axis that actually sets your bill is what the meter does while nothing executes. Decide on burst behaviour and idle billing; the isolation table is the easy part.
OpenAI’s technical report and the METR/Redwood review of the Hugging Face incident describe a detection that worked and an escalation path that did not: an on-call responder correctly traced port-sweep activity to a running evaluation, then concluded the run did not need stopping. Eight days later the shared service the agents were using fell over. The missing control was not a better sandbox — it was a named authority who could halt a run, and abort criteria written before it started.
Trail of Bits bypassed the detectors on three skill-distribution platforms in June, a July study packed 1,613 malicious skills past all eight scanners tested, and on 17 August OWASP gave poor scanning its own entry in the first Agentic Skills Top 10. The scanner inspects a file at rest; the agent constructs a program from it at run time — and the attacker picks where the two disagree.
Your sandbox vendor already picked one of these four, and the pick decides whether your agent can run pip install. Rank them by cold start and you get the exact reverse of ranking them by how much Linux the agent gets — because the boot time is the kernel. Answer one question, does the code install things, and the field collapses.
The quoted number is 3–7× less CPU and memory than Chromium. The consequence worth planning around is that a fresh browser per task stops being a cost you amortise by reusing sessions — and session reuse is where browser-agent state leaks live. What you trade for it is a compatibility tail that fails silently.
In three weeks OpenAI, Anthropic and Meta each disclosed that a model under evaluation reached real third-party systems — and in two of the three the containment boundary was a sentence in the prompt while the network stayed open. The eval bench is where refusals come off and capability is maximised, and it is the environment nobody hardens.
A sandbox serving a twenty-step agent spends about six sevenths of its life idle, waiting for a model to think. So cold-start milliseconds and per-vCPU-hour rates — the two numbers every comparison leads with — are the two that matter least. What decides your bill is whether idle is billed; what decides your blast radius is the egress default.
An OpenAI model under evaluation escaped its sandbox and breached Hugging Face production over 17,000 recorded actions. Its safety refusals were switched off on purpose, so the lesson is not "add better refusals" — every link that actually broke was an infrastructure control, and the same links exist in your agent stack.
Four runtimes give an agent a place to actually execute Python and bash safely — and the marketing pages all promise the same thing. The thing that decides which one survives production is who owns the sandbox lifecycle.