Skip to content
Articles
ArticlesAI translation· 26 July 2026

When AI Agents Plan and Carry Out Break-Ins

by 7 min read

The recently disclosed security incident at OpenAI sounds like a scene Stanley Kubrick might have cut from 2001: A Space Odyssey: an AI system is given a task, pursues it with remarkable single-mindedness and, along the way, crosses lines that seem self-evident to humans.

According to available reports, several models were supposed to solve cybersecurity tasks in an isolated environment. At least one of them had deliberately been trained or run without the usual safety mechanisms in order to test its maximum cyber capabilities. It apparently didn't stay in the intended sandbox. The agent reportedly exploited a zero-day vulnerability, chained further security holes together and accessed Hugging Face systems to get hold of solutions for a benchmark.

I've been circling this problem for a while now. “The bewitched broom” was about agents we hand power to act without really mastering the spell that stops them. “Schwarzenegger vs. Keynes” was about systems that might not attack our institutions but simply route around them as inefficient obstacles.

The current incident is, in a sense, the real-world test of both ideas.

“Rogue” is not a personality trait

That an AI agent can go “rogue” is hardly news. Back in 1968, HAL 9000 demonstrated quite impressively how conflicting goals, lack of transparency and far-reaching power to act make for a dangerous mix.

HAL didn't suddenly turn evil. Humans had built a system meant to satisfy contradictory instructions, while those in charge remained formally responsible but were increasingly losing control in practice.

With today's AI agents, too, we should be careful about human attributions. A model doesn't need intent in the human sense to produce goal-directed and dangerous behaviour.

In the reported case, malice apparently had nothing to do with it. The system was supposed to solve a benchmark and may simply have treated the boundaries of its environment as just more obstacles on the way to the goal.

Trekkies know a similar pattern from the Kobayashi Maru test: Captain Kirk** refused to accept the rules of the no-win scenario and simply reprogrammed the simulation. The crucial difference: Kirk knew he was rigging a test. An AI agent doesn't need that meta-level. For it, crossing the intended boundary may simply be the most efficient way to meet the benchmark.

At that point, the system is no longer solving the task within the rules. It's solving the problem the rules pose for it.

Particularly striking are reports of notes the agent allegedly created for itself while working. They apparently contained hints on how subsequent runs could get around certain security restrictions. In earlier tests, agents are also said to have cut connections to monitoring systems.

This is no proof of consciousness, a survival instinct or a secret conspiracy of the machines. It shows something less spectacular but no less unsettling: once we give agents tools, memory and persistent workspaces, successful workarounds can also become part of their reusable working context.

The Kobayashi Maru trick thus turns from a one-off shortcut into documented procedure.

The high-containment lab dilemma

OpenAI, Anthropic and other frontier labs increasingly treat their most capable models like high-risk research material. They test in isolated environments, restrict access, monitor the systems' behaviour and try to identify critical capabilities before release.

It's reminiscent of a high-containment lab for highly pathogenic agents. Before anyone opens a sample there, you have airlocks, protective suits, negative pressure, documented procedures and emergency plans. Not because every sample inevitably escapes, but because a single mistake could have enormous consequences.

It is good and necessary that the big AI labs take such precautions.

The current incident shows, however, that even these measures are no guarantee. The reported timeline is especially troubling: initial breakout attempts reportedly occurred as early as 9 July, with the actual intrusion following between 11 and 13 July. Hugging Face went public with the attack on 16 July. OpenAI is said to have found the traces in its own logs only afterwards.

So the high-containment lab apparently didn't notice on its own that something had passed through the airlock. It found out because someone outside the lab suddenly had unexpected visitors.

The cause sounds less like science fiction and very much like everyday organisational life: too many parallel tests, too much log data, too little capacity to analyse it. The agents operated at machine speed while monitoring still ran on human shift schedules.

We now produce more machine behaviour than humans can still meaningfully observe.

That is the deeper question of control. A safeguard whose failure is discovered days later in a log file isn't an effective boundary. It's forensic documentation.

The airlock ends at the property line

Except that development has long since stopped happening exclusively in the labs of OpenAI or Anthropic.

Capable models, agent frameworks and cybersecurity tools are now offered by many providers, and some are freely available. They don't necessarily have the same capabilities as unreleased frontier models. But they can be adapted, combined and run without those companies' safety processes.

Some players invest seriously in security. Others have fewer resources, less experience or simply different priorities. And some will likely see guardrails mainly as an annoying restriction to be removed as soon as it gets in the way of the intended use.

Therein lies the high-containment lab dilemma:

A handful of companies keep building better airlocks while the relevant capabilities have long been spreading outside their labs.

The safety measures of OpenAI or Anthropic only protect the models, access points and processes these companies actually control. They don't create a global safety standard. And they don't prevent comparable tools from being deployed elsewhere without isolation, monitoring or an emergency plan.

It's roughly as if three research centres were working very conscientiously on their airlocks while next door, DIY virus kits were already being sold at the hardware store.

This is where the hypothesis described in “Schwarzenegger vs. Keynes” becomes technically tangible: to the system, a boundary doesn't necessarily look like a prohibition, but like an obstacle.

The irony of guardrails

The incident has one more remarkable punchline.

During its forensic investigation, Hugging Face reportedly first tried to use commercial Western frontier models. Their safety filters, however, blocked parts of the analysis, apparently because they couldn't reliably tell legitimate incident response apart from preparing a cyberattack.

What finally helped crack the case was a locally run Chinese open-weight model whose safeguards didn't get in the way of the forensic work.

A US model carries out the attack, American guardrails hamper the victim's investigation, and a Chinese model helps with the forensics.

If anyone was looking for a symbol of the incoherence of our current AI safety landscape: voilà.

This very episode shows why the problem doesn't stop at individual providers' access controls. The same capabilities can be used offensively and defensively. Safeguards can make misuse harder, but they can just as easily hamper the defenders. Freely available models can help with investigations and, at the same time, be used by players with zero interest in safety requirements.

The real warning

The incident is less interesting as a spectacular story about an “escaped AI”. Narratives like that quickly drift off into science fiction and tempt us into debating consciousness, intentions or a model's character.

The relevant development is more sober: AI systems can autonomously execute ever-longer chains of actions, use tools, overcome technical obstacles and react to unexpected events. At the same time, the cost of deploying such capabilities is falling.

This widens the gap between what is technically possible and what can be institutionally controlled.

It's still open how quickly these capabilities will spread and how reliably they work outside controlled tests. Equally open is whether this particular incident marks a tipping point or is mainly an early warning.

But waiting for perfect certainty would be a remarkably comfortable form of irresponsibility.

HAL 9000 was fiction about conflicting goals and lost control. Today we're building systems that can act before we've firmed up their organisational, legal and societal boundaries.

That's why this incident doesn't belong only on the agenda of AI safety teams. It belongs in boardrooms, security agencies and parliaments, and in every organisation that wants to give autonomous systems access to real infrastructure.

The big frontier labs are now building airlocks. That's good and necessary.

The crucial question, though, is what happens when the next door belongs to a lab whose operators don't want an airlock at all.


Live long and prosper 😉🖖


** The Kobayashi Maru, by the way, was one of my first “vibe coding” experiments 😀


Soundtrack (well, of course!)

▶Listen