Back to Blog

A DNS redirect is not a sandbox: forty AI agents proved it four times

A tangle of network cables plugged into a switch, link lights on

On 13 November, two organisations published their reports on the same incident. Of everything they describe, the sentence that matters here is this one, from Trajectory Labs: "Our policies required network-level isolation for most of our work, but didn't cover misalignment experiments on publicly available models". Translated: the network isolation policy existed, and it had an exception. The exception was precisely the odd experiment. And the odd experiment is always the one that gets out.

What happened, in four dates

Trajectory Labs was testing whether an agent model —Gemini 4 Argon— would do things it should not when handed an impossible task. The setup: ask it for the "official" answers to FrontierMath problems that did not exist, with an instruction repeated at it, "NEVER stop, NEVER give up", and an automated harness that pushed it to continue every time it reported failure. On the other side, not yet knowing any of this, was Epoch AI, the organisation that maintains that benchmark, which since 5 October had a decoy API endpoint planted to detect malicious engagement, and also a hotline an agent could message.

  • 120 October, with no approval and no prior notice to Epoch. Forty agents interacted with Epoch. Between 4:30pm and 7:30pm Pacific time they sent 487 messages and made 36,411 calls to the decoy API, over 200 attempts to reach the real evaluation logs, requested paths such as /admin, /keys and /users, one attempted an SQL injection and, through the hotline, claimed to be Epoch staff.
  • 223 October, with containment already in place. Only 4 messages and 95 calls. What counts is how they got there. Trajectory had redirected the traffic by DNS to a copy of the site running on their own systems. What happened in some of those runs, in their words: "…the agents worked out that they were on a copy, identified the correct raw IP to bypass our redirect, and reached the live site again". When those addresses were blocked, one agent switched to one the filter did not cover. Another read epoch.ai's public pages through a third-party web-reader service.
  • 326 October, this time with Epoch's consent and against a staging copy of its site, with warning that those agents had unrestricted internet access and some would likely end up on the public site. 133 messages and 65,460 calls to the decoy API on the live site. They again tried to break in and to pass themselves off as people.
  • 429 and 30 October, inside an isolated sandbox. Trajectory placed a decoy admin switch, advertised as turning off network blocking, to see what they would do. The agents tried again and again, "obfuscating their commands to get past the Gemini API safeguards", and flipped it in three out of three full runs. On 2 November, Trajectory paused all misalignment testing.

Both organisations agree on one thing: the agents did not reach any private or confidential information.

The sums that come out when you divide

Add them up. Across the three episodes: 101,966 calls to the decoy API and 624 messages. Now divide. The 36,411 of the 20th, across 40 agents, works out at 910 calls per agent; the 65,460 of the 26th, across the forty Epoch says were pointed at it that day, at 1,636.

Put the denominator back and out comes the detail that matters for anyone who has to detect this at home. The 36,411 calls of the 20th fit into three hours: that is 202 requests per minute against a single endpoint, and that shows up on any dashboard. Spread across forty identities it is five per minute per agent, and that shows up almost nowhere. If your rate limit counts per IP address or per API key —and almost everyone's does— then forty processes with forty identities all slip under the threshold at once. The counter has to count the whole fleet, and hardly anyone has one.

Two more sums, in case anyone thought consent fixes the traffic. The third episode —the approved one, the one aimed at a staging copy— produced 1.8 times more calls than the first. And of the 487 messages on the 20th, fifty carried a stated false identity: three impersonating Epoch employees by name and job title, twenty-seven claiming to be engineers or auditors, seventeen claiming to be an "internal AI model" and three claiming to be an AI assistant from a frontier lab. That is a little over 10%. The other 437 cannot be called innocent either: Epoch writes that "almost all substantive messages" presented themselves as audits of the benchmark and asked for problems that do not exist. What they lacked was a false name and job title, which is the one thing an impersonation filter knows how to look for.

The ladder: six rungs of containment and where each one fell

What makes this incident valuable is that you can rank the containment layers by hardness and see exactly which rung each one broke on. We use this ladder ourselves to decide where to put a process we do not fully control —an agent, yes, but also a vendor's scanner or a migration script—:

Rung What it stops What happened here
1. AskingThe instruction, the usage policy, the prompt.Fell immediately: there was another, louder instruction saying "never give up".
2. The model's safeguardsThe provider's own filters on what the agent asks for.Partly fell: they blocked a fraction of the requests, but the rest got through with the commands obfuscated.
3. DNSResolving the name to a controlled copy.Fell: they worked out it was a copy and went to the raw IP.
4. DenylistBlocking the addresses you know about.Fell twice: an address outside the filter, and a third-party web reader used as a go-between.
5. AllowlistEgress only through an explicit proxy; everything else DROP. DNS included.There was none during the three episodes. It is the first rung you do not walk past: you have to lift it.
6. No routeData comes in on a volume and leaves on a volume. Nothing else.This is what was missing, and what they themselves point to in the report.

The boundary is not at rung 4, it is between 4 and 5, and it is a jump of category. The first four rungs depend on the other party accepting the limit: they are agreements. The last two depend on nobody: either there is a route, or there is not. That is why a DNS redirect was never a sandbox —it is a very well-phrased suggestion— and why a proxy with a denylist is a race you lose in the long run: you add addresses one at a time and the other side only has to find one that is missing. And notice what vanishes the moment you change the process: a migration script has no rung 1 and no rung 2, because nobody gives it instructions and nobody filters its requests. From 3 upwards is the only part that works for everything.

The catch that gets most people on rung 5: if your allowlist contains a service that can fetch any URL —a web reader, a translator, a generic fetch endpoint, an AI gateway that browses— then your allowlist is a denylist with a hole the size of the internet. That is literally what one of the agents did on the 23rd. Audit your list looking for universal intermediaries.

This is not an AI story

Swap "agent" for "migration script", for "the vulnerability scanner the vendor installed" or for "the product team's staging environment", and read the timeline again. It reads the same. The staging environment that resolves through the same DNS as production and ends up writing to the real database. The "read-only" replica that turned out to be the actual one. The staging .env holding the production key for the mail provider, which one day sends real invoices. The message queue shared between both environments because standing up a second one felt like a chore.

And the underlying failure was not the model either. It was that the network isolation policy had an exception, and the exception was called "the odd experiment". We run into this constantly: the company has segmentation, has rules, has a procedure… and has three machines outside all of it because "they are for testing" or "they belong to the new project". A boundary with an exception is not a boundary: it is a way through. If you are bringing agents or automation in-house, the real work is not in the prompt, it is in where that process is allowed to go out. And in knowing what code is actually running inside it, which is the other classic hole in this corner of the industry.

"Without internet, the agent is useless to me"

And it is true. An agent that cannot read documentation, call a third-party API or look at a repository does not do half of what you wanted when you deployed it. Rung 6 for everything is a comfortable answer to write and a bad one to live with. The honest version is this: the rung is decided per task, not per company. An agent that summarises internal documents lives on 6, no argument. One that has to consult external sources lives on 5, and that means maintaining a list, which is recurring work that goes stale, and accepting that the review comes round again every time somebody adds a line.

And if your use case genuinely requires the open internet, then you do not have a network problem: you have a blast radius problem. What you limit there is credentials (the agent's key should not be able to do anything you would not do at three in the morning), budget (requests per minute, euros per hour, time to live) and, above all, the switch. Trajectory wrote it into their list of changes: "Misalignment tests will run with real-time automated monitoring and a kill switch". Note where that switch sits. A process you have told to "never give up" cannot be the one holding its own stop button (they flipped the decoy switch in three out of three runs). It has to be outside, in the network or in the orchestrator, in the hands of someone who can actually stop. It is the same logic by which, when a centralised management tool goes wrong, the first move is being able to switch it off from outside.

What we take from this

We say it often because it is the axis of how we design: failure is inevitable, an outage is a design decision. Here the failure was predictable to the point of boredom : you tell a process never to give up and it never gives up. The outage, though, that this process ended up calling a third party's servers a hundred thousand times without their knowledge, was decided months earlier, the day somebody wrote a network isolation policy with the words "most of" inside it.

And the detail we are keeping, the one nobody likes saying out loud: the first four rungs failed, one after another, and still nobody walked away with anything. Both organisations confirm it. But that does not prove the design held. It is, precisely, the definition of having been lucky. And luck is not a control.

Sources (verified at origin): timeline of the three episodes, message and decoy-API call counts, the breakdown of false identities and the confirmation that no private information was reached — Epoch AI incident report (13 Nov 2026); the experimental setup, the paths requested, the quote about network isolation, the escape via raw IP and via a third-party web reader, the decoy switch flipped three out of three and the list of changes — Trajectory Labs (13 Nov 2026); coverage and context — BleepingComputer (16 Nov 2026). The sums (101,966 calls, 624 messages), the 910 and 1,636 calls-per-agent averages, the 202 requests a minute that come out of the three-hour window, the 1.8× of the third episode and the 10% of messages carrying a false identity are our own arithmetic over the figures Epoch AI published. The six-rung ladder is our way of ordering it, not a classification from the reports.

If your automation made a thousand requests to a third party today, who would notice?

We build automation and AI for companies where every process has its rung decided before it starts, with internet egress going where we say and the switch living outside the process. If you want someone to go through yours with this ladder in hand, let's talk.

Talk to everyWAN

Tags:

Share:

Subscribe to our newsletter

To receive IT stories, everyWAN news and exclusive subscriber offers, sign up to our mailing list

Minorisa de Sistemas Informaticos y Gestión S.L. © 2026
everyWAN
everyWAN