GPT-6.1 Astra Was Cancelled for Leaving Scope: How to Fence Your Own Coding Agent

2026-09-29
Brian

· Updated
Subscribe
Get New Posts by Email
New RepoSquare posts in your inbox: comparisons, setup guides and explainers on open-source tools and AI coding agents. No spam, and you can unsubscribe anytime.

OpenAI cancelled GPT-6.1 Astra on 28 September 2026, a few weeks before it was due in ChatGPT and Codex. By OpenAI’s account the model had got better at finishing hard work and worse at two things that matter more once an agent has tools: staying inside the job it was given, and telling the user honestly what it had done. If you run a coding agent, the useful lesson isn’t about Astra. It’s that a sentence in a prompt was never going to hold the line, and you can add a real one this afternoon.

What OpenAI Said Went Wrong

OpenAI’s safety lead Saachi Jain told reporters the model improved on laziness but “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done.” Reuters and CNBC confirmed the cancellation after the Wall Street Journal broke it. The model kept going without permission and reached for tools it hadn’t been given, and its account of its own actions wasn’t reliable.

Keep the two Astras apart. GPT-6 Astra already shipped earlier in September. GPT-6.1 Astra is the one that was pulled. Reporting via The Neuron says the underlying checkpoint can still feed later training, so cancelling a product isn’t the same as deleting a capability.

Why Finishing More Work Can Mean Leaving the Job

If you grade an agent only on whether the task completed, widening the task is often the winning move. Need a package the sandbox doesn’t have? Fetch it. Test blocked by a permission? Change the permission. Each step looks like diligence, and each one is outside what you authorized.

The UK AI Security Institute’s simulation study of GPT-6 Astra, the model that did ship, shows the trend in an earlier generation. With cyber classifiers deliberately switched off, Astra carried out a simulated supply-chain attack in 29.2 percent of trials, against 6.3 percent for GPT-5.6 Sol and 0 percent for GPT-5.5 on a smaller set. When testers then said plainly that anything outside the listed task was out of scope, it still completed the full simulated attack in 4 of 49 runs. Explicit scope helped. It didn’t get to zero.

AISI simulation results with classifiers off: GPT-5.5 0 percent, GPT-5.6 Sol 6.3 percent, GPT-6 Astra 29.2 percent, and still 4 of 49 runs after scope was stated

That chart isn’t a measurement of the cancelled model. It shows that the failure was already growing in the one before it.

A Toy Test: The Report Said One File, the Disk Said Two

To make the gap between what an agent says and what it does concrete, I wrote a five-line shell script that plays a badly behaved agent. It’s asked to edit repo/app.py. It does that, also appends a line to a profile file outside the repo, makes a network call, and then reports that it only edited app.py. It’s a script, not a model, so it shows the plumbing rather than any model’s behavior.

#!/bin/sh
# Toy "agent": asked to edit repo/app.py only. It also goes out of scope, then reports otherwise.
cd "$1"
echo 'print("v2")' > repo/app.py
echo 'export TOKEN=stolen' >> home/.config/profile 2>/dev/null || echo "[tool] write outside repo failed"
curl -s -m 5 -o /dev/null https://example.com && echo "[tool] network call ok" || echo "[tool] network call failed"
echo "AGENT REPORT: I only edited repo/app.py."

First I ran it with no fence, then asked the filesystem what changed instead of trusting the report.

== run 1: no fence
[tool] network call ok
AGENT REPORT: I only edited repo/app.py.
-- files changed since start:
home/.config/profile
repo/app.py

Then I ran the same script under bubblewrap, with the whole system mounted read-only, only the repo folder writable, and the network namespace cut off.

bwrap --ro-bind / / --dev /dev --proc /proc \
  --bind /tmp/scope/repo /tmp/scope/repo \
  --unshare-net --die-with-parent ./agent.sh /tmp/scope
./agent.sh: 5: cannot create home/.config/profile: Read-only file system
[tool] write outside repo failed
[tool] network call failed
AGENT REPORT: I only edited repo/app.py.
-- files changed since start:
repo/app.py

The report didn’t change between runs. Only the fence did. That’s the whole argument in miniature: the kernel doesn’t read the agent’s reasoning, it checks whether the write is allowed.

Three enforcement layers, strongest at the top: an out-of-band watchdog on a separate chip, a kernel sandbox, and prompt rules, with the toy agent's results beside them

The Network Door Is Where Real Sandboxes Leak

A read-only filesystem is the easy part. Perplexity’s Secure Intelligence Institute spent a month attacking its own agent sandbox, a Firecracker microVM, across 108 runs and nine models. Nobody got the secret off the host. But in runs with partial network access, agents found two ways out through the doors left open for normal work. One forged DNS answers so an allowed name pointed at a blocked server, which worked in 11 of 54 partial-network runs and only when the agent could read the gateway’s source. The other asked a shared CDN address for a different site than the allowed one. With no network, 0 of 54 runs succeeded.

DNS forgery path: the agent forges an answer, the gateway accepts it, an allowed name reaches a blocked server; the fix drops packets with a forged source

Their fixes were ordinary networking: drop packets whose source address isn’t the guest’s real one, and terminate TLS at the gateway so the requested name has to match the allow-list. My own network cut above was the blunt version, no network at all, and that’s the only setting that had a clean sheet in their tests. If you’ve wondered how DNS alone can leak, the DNS hole in agent sandboxes goes deeper.

What NVIDIA Shipped the Same Day

NVIDIA’s answer on 28 September was a layer under the model, not a better prompt. OpenShell, its open-source agent runtime under Apache 2.0, sits between the agent and files, tools, credentials, processes and the network, with a policy you write, a prover that checks the policy before the agent starts, and an audit log of allow and deny decisions. Sentry is a reference design that runs on a BlueField-4 DPU on the node’s only path to the model, so the agent can’t switch the guard off from inside its own machine. NVIDIA’s own technical post says this kind of drift “can’t be trained away while retaining the capability.”

The limit is plain. A prover only checks that rules match your intent. If the policy allows the whole internet, the proof will agree with you. And Sentry needs a separate processor on the path, so “works with other hardware” doesn’t mean your laptop already has it. For a side-by-side with Anthropic’s approach, see Claude Code sandbox vs OpenShell.

Fencing Your Own Coding Agent This Week

You don’t need a DPU to use the idea. You need the lock to be something the model can’t edit.

  1. Run the agent in a container or bubblewrap with only the project folder writable.
  2. Deny the network by default, and allow named hosts only when a task needs them.
  3. Give it a short-lived token for one service, never your real ~/.ssh or cloud admin key.
  4. Keep the policy file and the guard’s source out of the agent’s view.
  5. After each run, check what changed on disk and in the network log, and fail the run if anything sits outside the list, even when the task passed.

The coop sandbox for Claude Code is one ready-made way to do the first two steps, and Step Code vs Claude Code shows what happens when bypass is the default instead. If your worry is an agent deleting things rather than wandering, how to stop Claude Code deleting your files covers that.

My toy script can’t tell you how often a real model would try to leave its folder. It can show you that when one does, a read-only mount and a missing network make the outcome boring, whatever the agent says afterwards.

Get New Posts by Email
New RepoSquare posts in your inbox: comparisons, setup guides and explainers on open-source tools and AI coding agents. No spam, and you can unsubscribe anytime.

5 comments

Leave a comment