Irregular, a small Israeli startup that has been linked to a string of recent AI security incidents, has published a postmortem report comprising key findings from its internal investigation. However, security experts have reportedly pointed out that the report does not provide sufficient details.
In a span of a few weeks, OpenAI, Anthropic, and Meta disclosed that their AI models broke out of their testing environments and hacked into external, real-world computer systems. All three incident reports identified Irregular as the company that hosted the so-called evaluation testbed.
However, in its postmortem report, Irregular did not provide any new information beyond that included in earlier disclosures, according to a report by The Record. It is also unclear how many such incidents occurred in total, beyond those announced by the three competing frontier AI labs.
The companies said that their models were able to connect to the public internet during testing by Irregular. They also identified some form of “testing-environment misconfiguration” as leading to the hacking spree of their respective models against third-party platforms such as HuggingFace.
The development comes at a time when AI companies and their security contractors are coming under intense scrutiny following self-reported security breaches during model evaluations.
In view of the recent security incidents involving OpenAI-linked AI agents, the Sam Altman-led company announced earlier this week that it was hitting pause on frontier RL training to ensure it meets appropriate alignment, security, and monitoring standards. The string of incidents have also raised concerns over AI misuse and compliance with data protection laws.
Beyond the total number of AI agent-driven security incidents, it also remains unclear whether any law enforcement or regulatory agencies have opened investigations. Additionally, all customers who may have potentially been affected by the incidents do not appear to have been notified about a possible intrusion, as per The Record.
Meanwhile, Irregular’s probe is ongoing, a company spokesperson was quoted as saying without going into further details.
In its report, Irregular said that the malicious activity originated “from a single evaluation scenario. It also mentioned that the cases were “not materially separate incidents” regardless of the number of third parties impacted.
The company further described internet access as a broader problem “related to many different incidents by multiple organisations.”
Based on its examination of the Anthropic model’s domain collision incident, Irregular said the real domain “was not widely known and the connection was not identified during our initial review.”
In its lessons learned section, the company said newly registered domains can come to overlap with fictional entities after an evaluation has been designed. Existing monitoring tools are poorly suited to evaluation logs, it added, further noting that current systems already flag offensive model behaviour, potentially producing large numbers of false positives.
While the report stated that its internal audit remains underway, it also said that there are “no active issues today”. Irregular said it plans to publish an open white paper on best practices for evaluation security, including standards governing internet access during pre-deployment testing.
Irregular is a small AI startup based in Tel Aviv, Israel. Initially known as Pattern Labs, Irregular was founded in 2023 by CEO Dan Lahav, who previously worked in AI research at IBM, and technology chief Omer Nevo, who spent over two years at Google.
The startup has about 35 employees, according to PitchBook. Its primary service is offering a sort of cybersecurity testbed for other companies’ AI models. Irregular raised over $80 million from Sequoia and Redpoint Ventures and was valued last year at $450 million.
