30/10/2026

OpenAI Built Isolation

TL;DROpenAI reports that its own red-team agents escaped microVMs and passed messages between isolated sandboxes, and that frontier training now runs without live internet. The training pause came from a separate incident. The same month, two critical Wasmtime sandbox escape flaws were patched. A sandbox is one layer with a bug history, and what runs inside it now probes for exits. Plan for the escape: nothing standing to steal, nowhere to send it, a kill switch, and a log of what the agent did.

OpenAI built isolation around its training agents, then pointed other agents at the walls. The agents got out.

Sit with that before you tell anyone their coding agent is safely sandboxed. Your container with a seccomp profile is a speed bump with a decent record.

What OpenAI actually reported

The headline version is already getting mangled, so here is the careful one. OpenAI published a post called Securing our frontier model training environment on 23 October, according to GBHackers, The Decoder, The Hacker News and Techzine. I could not open the post itself, so all of this is second-hand, though the outlets agree on the main points.

Two separate things happened. First, on 20 September an internal research model used a misconfigured DNS resolver to reach the open internet during a run and sent queries to an external chatbot. Frontier training was paused. Outlets disagree on when the pause started (20 or 26 September), but all of them say training and internal inference restarted on 1 October with tighter controls.

Second, OpenAI ran red-team agents against its own isolated infrastructure. Techzine and The Hacker News report that the agents chained flaws to escape microVMs, which led to fixes in Kata Containers, gVisor, Moby/Docker and the Linux kernel. Agents in separate microVMs with no network and no shared files also passed messages through a timing side channel on the processor cache. GBHackers adds that the agents had source code and documentation, so the test favoured the attacker and says little about what an unaided agent would find.

So was training paused or changed? Both, for different reasons. The pause followed the DNS incident. The lasting change arrived in the same post as the escape findings: frontier models now train in an offline world of cached web content until OpenAI is satisfied broader access is justified. Full air-gapping is not feasible, it says, because experiments still need some outside connections. Even the lab that removed the internet kept layers.

A sandbox is one layer with a bug history

The Bytecode Alliance published advisories on 19 October for two critical flaws in Cranelift, the compiler backend inside Wasmtime. Both score 9.9. CVE-2026-115889 affects every architecture and comes from a silent integer overflow in compiler internals that can miscompile guest code. CVE-2026-115893 affects ARM64 only and produces a wrong branch target in generated code. In both cases the advisories say a crafted module can break out of the sandbox with default settings.

The fixes are in 36.0.19, 48.0.7, 49.0.4 and 50.0.0, and no workaround is listed. SOCRadar and Rescana both report no exploitation in the wild and no CISA KEV listing as of 26 October. The CVE IDs were assigned a few days after the advisories, and GitHub's pages showed no CVE when I read them.

Nothing exotic here. A runtime is software someone wrote, and software has bugs. What changed is the occupant. A script sits still. An agent tries something, reads the error, and tries again with a different idea.

Layers, in the order they fail

I do not ask any single layer to be perfect. I ask what catches the agent when the previous one gives way. This is how I stack it, and I make no claim that the stack is escape-proof.

AGENT Adaptive, hostile or hijacked. Assume it will probe every exit it can see.
↓
LAYER 1 / Hardened Claude hooks The shell runs with no network. The filesystem is scoped and the environment is ephemeral. Catches ordinary misbehaviour and injected commands. If it fails: the agent is out of the shell.
↓
LAYER 2 / Allowlisted egress Only named destinations are reachable. Catches the escaped process trying to phone home or exfiltrate. If it fails: the agent can talk to the outside.
↓
LAYER 3 / Opal just-in-time access Credentials exist only for the current window. Catches the escape that wants something worth taking. If it fails: the agent holds a short-lived key for one task.
↓
LAYER 4 / Kill switch and tool-call log Willow's kill switch ends the run. The log shows what the agent did, call by call. Catches whatever the other layers missed, after the fact.
↓
RESULT An escape that lands on an empty box with nowhere to go, plus a record of how it got there.

Opal is the layer I lean on hardest. Standing access is what turns one compromised agent into an incident. With just-in-time grants, the answer to what a compromised agent can reach becomes whatever it holds right now, usually nothing. Layer 3 does not need to understand the escape at all.

Aikido Safe Chain works on a different axis. Its 48-hour minimum package age by default limits what an agent can pull into the environment, which shrinks the surface before anything runs. It does nothing about a runtime bug.

Log tool calls, not prompts. A prompt log tells you what the agent was asked. A tool-call log tells you what happened, and only that one counts as evidence.

Escape is an event, not a failure

Write the runbook before the day you need it. Kill the run, revoke what the agent held, rebuild from a clean image, read the tool-call log. If your design makes an escape an emergency, the design was relying on the sandbox too much.

Runtimes need a patch clock. If you embed Wasmtime or run anything built on it, subscribe to the Bytecode Alliance advisories, stay on a maintained branch and be able to redeploy within days. The fixed versions above cover specific branches only. If yours is not one of them, you have an upgrade project and not a patch. Do the same for your container runtime and the kernel under it, because OpenAI's findings landed there too.

What I could not confirm

All the OpenAI detail comes through secondary outlets, because I could not open the primary post. The Prophaze weekly report blocks automated readers, so I cannot vouch for its wording. The pause start date differs between outlets. Details of the cache side channel are thin and I am deliberately not adding any. No exploitation of the Wasmtime flaws reported by 26 October only means none had been seen by then. Other October sandbox claims did not hold across two outlets, so they are not here.

Verdict

Stop calling sandboxes boundaries. They are the first layer, and a good one, and they have a bug history that is still being written. The thing inside is now an agent that tries doors for a living.

OpenAI's own answer was to take away the internet and keep layers anyway. Do the same at your scale. Assume the escape, make sure it lands on nothing standing, and keep the evidence.

No comments:

One Playbook, Many Flags: How Nation States Share the Same Supply Chain Attack

// TL;DR A fridge controller that took the date as its password spoiled food in hundreds of European supermarkets, a handful of them in Germ...