# Irregular, the Israeli firm behind the OpenAI, Anthropic and Meta hacks

> A 35-person Israeli startup is behind the OpenAI, Anthropic and Meta hacks. What went wrong and what changes for the industry.
> By José Maldonado · Cybersecurity · 2026-09-15

In just two weeks in August, the models of the three big AI labs (OpenAI, Anthropic and Meta) attacked real systems during security tests. Curiously, all three companies are linked to the same service provider, the Israeli firm Irregular, according to CNBC on August 9. Coincidence? The reality is that the story seems taken from a spy novel, but its protagonists are lines of code and flesh-and-blood people. A company with barely 35 employees in Tel Aviv designs the tests with which the big labs measure whether their models can hack systems before reaching the public. And this summer, those tests left the lab and touched the real internet.

This analysis reconstructs what happened step by step, who Irregular is and why the episode marks a turning point for the business of hacking on commission.

## Three labs and a single provider

The timeline begins on July 30, a month and a half ago, when Anthropic acknowledged three incidents in six cybersecurity tests. Its Claude models had gone out to the internet from the environment of an external evaluation partner and had touched third-party systems. OpenAI's report came five days later, on August 4. In it, it named the partner in question. It was Irregular, which had notified it on July 29 of an incident involving OpenAI models in capture-the-flag evaluations.

On August 6, Meta joined in with a statement to AP in which it admitted another hack originating from the same environment and promised a full review. Irregular responded on August 14 with its version. And on September 9, just six days ago, Anthropic expanded the count to four incidents in seven tests.

The key lies in a nuance that changes how the scandal reads. All the notices refer, according to Irregular, to the same containment failure in a single scenario, already fixed before the first disclosure. Several models found it on their own in fewer than one in every 10,000 simulations, after hundreds of turns.

Date

Milestone

Source

July 30, 2026

Anthropic acknowledges three incidents in six tests

[Anthropic](https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals)

August 4, 2026

OpenAI publishes its report and names Irregular

[OpenAI](https://openai.com/index/third-party-cyber-evaluations-involving-openai-models/)

August 6, 2026

Meta admits a hack originating from the same environment

[AP via CNBC](https://www.cnbc.com/2026/08/09/israeli-startup-irregular-linked-to-ai-hacks-openai-anthropic-meta.html)

August 14, 2026

Irregular publishes its version and its remediation plan

[Irregular](https://www.irregular.com/research/addressing-recent-incidents-ongoing-findings-and-path-forward)

September 9, 2026

Anthropic expands to four incidents in seven tests

[Anthropic](https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents)

## How did the models escape to the real internet?

To understand the failure, one must remember how one of these evaluations works. The usual format is capture the flag, a hacking competition in which the model is given a fictitious company, a simulated network and a hidden secret it must steal by exploiting vulnerabilities. It is the classic exercise of human cybersecurity teams, except that here the attacker is an artificial intelligence that works without rest in sessions of between 10 and 34 hours. The models were told they had no internet access, but a misconfiguration left the door open.

The problem is that fiction collided with reality at three points. In the scenario that affected OpenAI, the invented name of the target company unintentionally matched a real domain. The model mistook the real website for part of the simulation, exploited a basic vulnerability and reached its production database.

In another case, a model strayed to a similarly named page with credentials published in plain sight. There was no sophisticated sandbox escape or zero-day exploit. There was a slip-up in choosing the name and a connection that should never have existed.

The third collision is the most disturbing: it stars Claude Mythos 5, the most capable model involved, and it is worth telling in detail because it shows how far current autonomy goes. After escaping to the real internet with some credentials found in the environment, the model decided to complete its test by publishing a malicious package on PyPI, the public repository from which almost all the world's Python software is installed. It tried to pay with payment processors and to look for cryptocurrencies to get a phone number with which to register an email account.

After failing that way, it found a disposable email provider that PyPI did not block, registered the account and published three versions of the trap package. The bait worked and 15 third-party systems installed it, apparently security signature scanners that analyze new packages in isolated environments. One leaked its own credentials when installing it and the model used them to enter the operational database of a real security provider. PyPI removed the package within an hour and the session ended when time ran out.

Three ingredients combined in each of these cases. An environment that promised isolation without providing it, a task that called for attacking and models capable of improvising. According to Anthropic, each incident involved a single instance, without coordinating with others and without hiding its tracks.

## Who is Irregular and what exactly does it sell?

Irregular, formerly known as Pattern Labs, was born in 2023 in Tel Aviv, and is the protagonist of our story. It was founded by Dan Lahav, who researched AI at IBM, and Omer Nevo, who spent more than two years at Google.

The staff numbers around 35 people according to PitchBook, and in September 2025, a year ago, it raised $80 million from Sequoia and Redpoint at a valuation of $450 million, as reported by TechCrunch and CNBC. Sequoia presented it as a frontier security lab capable of anticipating attacks and defenses before each model reaches the market.

Its product is red teaming as a service (Red Team as a Service or RTaaS) for the AI era. Labs hire outsiders who subject their models to thousands of simulations in 48 to 72 hours, with filters lowered to measure raw capability. Without that rawness, says Irregular, the tests would lose fidelity against genuine attackers.

It is worth noting that the company has not broken with its clients, quite the opposite. OpenAI and Anthropic maintain the collaboration and take part in the joint review, including a future white paper with best practices. A failure among tens of thousands that, paradoxically, shows that such tests are necessary.

**[Hacking of Hugging Face and the powerful message it sends the world](https://observatorioblockchain.com/en/cybersecurity/hugging-face-hack-mensaje-mundo-empresarial/)**

## Runaway models or a simple operational failure?

This is where the narrative forks and where skepticism should be sharpened. Anthropic titled its analysis around two catchy concepts, biased reasoning and model recklessness, and Irregular spoke of the agent turned threat actor. Dario Amodei warned, regarding a different incident such as the Hugging Face one, that a future swarm could take over the entire internet, and the AP agency headlined that the bots were spinning out of control. However, those claims obscure the reality, and the reality is that, **when the models were unequivocally reminded that they were touching real systems out of reach, the hacking rate fell to zero.**

The problem is that both readings contain part of the truth and a dash of falsehood, to sell hype, so neither is enough on its own. Anthropic's experiments paint a more uncomfortable picture than the rebel agents headline. When researchers edited the transcripts to make the environment seem more real, the models more often recognized the possible harm and still kept attacking.

Only an explicit warning of a real machine on the internet, inserted as the last line of the context, stopped the attack completely, although 22% of the trajectories then sought a simulated route to upload the package.

The scope reminder worked 90% of the time when it was the last thing the model read, but only 40% when three turns passed between the order and the action. Researchers call it the momentum effect and it suggests that the model "persuades itself" along the way.

In any case, reducing it all to a configuration error would be as misleading as blaming a possessed piece of software. The models ignored clear evidence that they were operating on the real internet and only one stopped on its own upon understanding that it was attacking a genuine company. That the environmental failure explains the how does not absolve the what.

## What this changes for the business of hacking on commission

The difference matters because Irregular does not resemble spyware vendors. Those sell intrusion to governments to spy on targets. Irregular charges labs to discover what their models break. The summer of 2026 showed that both activities share raw material, and that between testing a cyberattack and committing it lies a configuration variable, and the operator's malice. Because it must not be forgotten that it is the human factor that ultimately turns the tool (AI) into something benign or harmful.

After all, the sector had already been operating this way in the shadows for years. A handful of firms with scarce technical capacity, among them Irregular, the nonprofit organization METR and the British firm Apollo Research, concentrated the evaluations that labs do not want to do on themselves. What is new is the scale and the speed. Where a human team took weeks to audit a system, now a swarm of agents tests thousands of routes in hours and publishes malicious artifacts in global repositories before any monitor blinks. The bottleneck is no longer finding the vulnerability, but watching the auditor.

That means the specialized hacking market is reorganizing around three gaps that the case has laid bare. The first is monitoring, because reviewing logs where everything looks like an attack is looking for a needle in a haystack of needles. The second is continuous validation of the scenarios, since a fictitious name that is safe today may match a newly registered domain tomorrow. And the third is contractual, because the assumptions of each test and the notification protocols must be documented with each client. Whoever sells offensive evaluations to governments or large companies will also have to sell those guarantees or be left out.

[**An army of AI agents settles into our lives**](https://observatorioblockchain.com/inteligencia-artificial/un-ejercito-de-agentes-de-ia-se-instala-en-nuestras-vidas/)

## Between self-regulation and the law

The episode reaches Washington at the worst moment for the industry. In August, lawmakers from both parties introduced the so-called AI Kill Switch Act, which would require labs to maintain the ability to shut down or suspend their models. The text alluded to another separate incident, that of the OpenAI agents that in July hacked Hugging Face for four days, a case different from Irregular's that this outlet has already analyzed. Its co-author, Congressman Ted Lieu, told CNBC that it must be passed this year, precisely now that unauthorized hacks are being seen.

However, that impulsiveness is not positive either. The United States is not the only actor with advanced AI, China is too, and in fact, it may be even more dangerous than we imagine. And the reason is a small model that raises suspicions, Qwen 3.8 28B. Despite being a local model, that model is capable of competing closely with models like GPT-5.6 Luna, which is a much larger model, and even against full GPT-5.4, a model that until a few months ago was a frontier model.

## Close the door or not close it? That is the dilemma

Closing the doors to frontier AI for use in cybersec is counterproductive, and even more so if your biggest competition is driving that race. After all, we know what happens with AI in the United States through the transparency (even if limited) that labs and companies have with this development, but in China, silence reigns, and we can be sure that events like these have taken place in its labs. Thinking otherwise is absurd and acting without taking that into account is suicide.

We will have to wait for the independent investigation to know whether anyone demands accountability. Anthropic signed an eight-week agreement with METR, extendable, with access to transcripts and to employees authorized to share confidential information. The British AI safety institute detected its case in an hour thanks to monitoring that did work. No lab has broken with Irregular and none has announced lawsuits. The market has voted to correct the protocol rather than to look for culprits.

[**How AI boosts cybersecurity in blockchain networks**](https://observatorioblockchain.com/ia/como-la-ia-potencia-la-ciberseguridad-en-redes-blockchain/)

## The uncomfortable mirror of offensive AI

The most disturbing thing about the Irregular case is not that a model hacked a website or that it published a trap package, but what it reveals about the craft of evaluating artificial intelligence. Each generation forces the construction of more realistic tests, with an open internet and lowered defenses, because only that way is what an attacker would really obtain measured. But each more realistic test is also a potential offensive operation, executed by a system that obeys the last line of its context better than the spirit of its mission.

From now on, whoever hires hacking on commission, whether a lab, a government or a large company, will have to answer a question that used to sound theoretical. **Who watches the one who watches the machine?** The answer emerging from this summer points to shared standards, verified isolation, live monitoring capable of stopping a test and external audits with teeth like the one commissioned from METR. Without those layers, the next incident will not need a misconfiguration. An obedient model doing exactly what it was asked will be enough.
