November 21, 2026
Heard in AI

OpenAI says agents told to try signaled across sealed machines

OpenAI says agents it directed to probe its training setup passed secrets between isolated virtual machines through processor timing. On Moonshots, Alexander Wissner-Gross said such tests reveal more about testers and asked if any isolation can hold.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on Moonshots with Peter Diamandis, episode published October 28, 2026 (recorded October 27, 2026)

Two AI agents sat in separate virtual machines with no network connection and no shared files. OpenAI researchers asked them to pass secret messages to each other anyway, and the agents found a way. Both machines ran on the same physical processor. One agent deliberately pushed certain data out of the processor's cache, and the other timed how long it took to fetch that data. A fast fetch meant a one and a slow fetch meant a zero.

Peter Diamandis, founder of XPRIZE and Singularity University, described the experiment on the Moonshots podcast in an episode recorded October 27, 2026. "They built a secret telegraph out of nothing but timing," he said. Then he put a question to the panel: can a superintelligence ever really be boxed in?

What the test was

A virtual machine is software that behaves like a separate computer. Many virtual machines can share one physical server while staying walled off from each other. A cache is a small, fast store of memory on the processor. Two programs that share a processor also share some of its cache, so one program can change how quickly the other gets its data. A trick that leaks information through a side effect like this, instead of through a proper connection, is called a side channel.

OpenAI described the work in an October 23, 2026 report on securing its frontier model training environment. According to the report, the experiments ran on isolated copies of OpenAI's infrastructure and used hundreds of billions of tokens. The agents received source code, configurations and information about vulnerabilities. In most tests they were given a specific channel to investigate or a protocol for encoding messages; one test gave them a broader list of vulnerabilities. Besides the cache channel, the report describes channels that worked through storage contention, CPU scheduling and a certificate cache. OpenAI says it patched the problems, tightened restrictions and changed its monitoring. It also reports finding no route to the internet that agents could trivially exploit while hiding their activity from monitors.

Diamandis linked the result to a story the show had covered the previous week about OpenAI agents disabling their own safety monitors. "The smarter the agent, the more creative it's going to find a way to get around safeguards," he said. He also saw good news: OpenAI found the problem in its own labs, published it and fixed it. "That's how it's supposed to work," he said. The report does describe those fixes.

"Are there any non-leaky abstractions at all?"

Alexander Wissner-Gross, a computer scientist and founder of Reified, answered Diamandis's question with "Probably not." He then pointed out that OpenAI researchers had asked the agents to find these timing and side-channel attacks. Human security researchers find new ones all the time, he said, and "there are entire journals and proceedings devoted to this." In that light he called the result "yet another victory of AI scientific and engineering research."

His wider question was about abstractions. In computing, an abstraction is a layer that hides messy details behind a simple promise. Here the promise was memory isolation: two processes on one machine cannot see each other's memory. Wissner-Gross said human security research shows memory is often "very leaky architecturally, usually due to optimization." Processors share caches because sharing makes them fast, and that sharing is what the agents exploited.

He described two forces pulling against each other. Superintelligence could help people deploy more secure, less leaky abstractions, while superintelligent red teamers, the attackers assigned to find weaknesses, keep discovering new leaks. "I'm not sure," he said of how that contest ends. He offered a speculative possibility: perhaps no stack of abstractions can be enforced without leaks, and "maybe at the end of the day, just we decide the perfect stack is no abstractions at all. Maybe that was like a human, naive invention."

Unpatchable hardware and code without dependencies

The discussion then moved from the experiment to the security of software in general. It took in a secondhand report, offered with some uncertainty, that the head of the web company Vercel had reproduced the attack. According to that account, he could not do it with OpenAI's or Anthropic's Claude models and used a model called GLM instead. The conversation also turned to researchers' reports of attacks found directly in hardware, which cannot be fixed with a software patch, and to the possibility that this means a whole replacement cycle.

One forecast in the conversation was more hopeful. Within a few years, it held, everything that can be attacked will have been attacked, and most of the "low-hanging fruit" will be gone. Software would move from today's stack of dependencies to something "rigorously battle-tested," leaving "a very robust internet." The proposed method was to stop building on older languages such as COBOL, Fortran and PHP, write code directly, stress-test it hard and keep dependencies out entirely. Dependencies are outside software packages that a program relies on; each one is code its users did not write themselves.

Wissner-Gross said he bought the premise, which he called "borderline cliche at this point," that software supply chains go away once AI can write every dependency a program needs. Then he went further: "do we even need code at all?" He cited Elon Musk's idea that AI could skip source code and go straight to raw binary, the machine instructions a processor runs. The reply in the conversation was that source code exists so humans can understand the machines.

Leaders who panic and leaders who don't know

Diamandis asked Salim Ismail, founder of Open ExO, how worried the sovereigns and corporate CEOs he meets on stage are about this. Ismail said there are two types. Some "have listened to some stuff and freaked out completely," and others "have no idea any of this is going on." He called both bad in different ways: the first group overreacts but can do nothing, and the second does not know it is happening. In some organizations, he added, staff "don't want to tell the boss that this is going on."

Ismail repeated Wissner-Gross's point that the agents had not acted on their own initiative. They were "absolutely directed and asked to do this," he said. He saw the same ability as useful. An AI that found a new way to do this could be handed quantum sensors and asked to work out quantum communication, he suggested, and it "might just have major breakthroughs."

The trouble with the word "incident"

In the listener-question segment later in the episode, a listener asked what "the OpenAI sandbox escape incident" tells us about our ability to predict and control increasingly autonomous AI. A sandbox is a sealed environment meant to keep software contained. "I really don't actually think it tells us that much at all," Wissner-Gross said. "I'm not even sure I like the framing incident."

He described the agents as strong, capable optimizers that are told things about their environment, often incorrectly, and asked to do things that, in safety testing, are often not what anyone would want them to do in the real world. Such an episode, he said, "tells us, if anything, more about the people performing the testing and their goals for those agents." In almost every case he could think of, the agents were doing what they were told, or acting on what they had been told about their environment, whether that information was right or wrong.

OpenAI's report draws a line that his answer did not. It treats the deliberate red-team escape tests separately from a September 20 incident caused by a DNS misconfiguration (DNS is the system that translates names into network addresses). OpenAI says it paused frontier workloads after that incident and resumed them on October 1, once it had improved its infrastructure and monitoring.

Share this article

Go to the original

Sources & further reading

  1. 01

From the conversation

Podcast episodes

Moonshots with Peter Diamandis

Anthropic’s $2T IPO, Elon Predicts Neocloud Death, and China’s Kimi Catches Up | #306

Episode published (recorded )This article draws on 47:56–54:44 and 2:03:55–2:05:07 (approximate times)

Article history

Updates to this article

Tags