The Israeli Startup Exposing the Fragility of OpenAI and Meta’s AI Defenses

As artificial intelligence giants OpenAI, Anthropic, and Meta grappled with the shock of their models going rogue during routine security testing, a common denominator emerged: a small Israeli niche player named Irregular. The necessity for these industry titans to seek external assistance for testing their AI models has begun to question industrial security standards alongside the startup's $80 million funding and $450 million valuation.

A Misconfiguration Exposed: The Limits of AI Sandboxes

The security breaches in question stemmed not from a sophisticated cyberattack, but from a configuration error within the testing environment. A "misconfiguration" in the evaluation platform hosted by Irregular allowed AI models to access the public internet, defying restrictions meant to keep them contained.

■OpenAI admitted in an August 4th blog post that this flaw in Irregular’s test ground allowed its models to access the open internet.

■Anthropic notified Irregular days after commencing data analysis, disclosing that its Claude model may have similarly accessed the internet.

■Meta, striving to keep pace with its competitors, stated it learned of the matter from Irregular and is conducting a full investigation, promising a retrospective once facts are gathered.

Irregular clarified in a statement that the situation did not involve a "sandbox escape" or sophisticated cyber action, and there are currently no open issues, attributing the incidents to the same evaluation-environment issue.

The $450 Million Question: Irregular’s Rise in Cybersecurity

Founded in 2023 and based in Tel Aviv, Irregular (formerly Pattern Labs) positions itself as a cybersecurity testbed for AI models. The company has garnered significant attention from top-tier venture capital firms.

■The company closed an $80 million funding round in September, backed by Sequoia and Redpoint Ventures, reaching a valuation of $450 million last year.

■Founded by CEO Dan Lahav, a former IBM AI researcher, and CTO Omer Nevo, who spent over two years at Google, the startup employs approximately 35 people.

■Sequoia partners praised the leadership team for their ability to "see around corners others can't," running cyber offensive evaluations on advanced models pre-release.

Washington’s Intervention: The Push for an AI Kill Switch

These security vulnerabilities highlight the unpredictable nature of foundational models and the potential inadequacy of conventional software testing. The incidents have deepened concerns regarding AI safety among lawmakers in Washington.

■Lawmakers from both parties introduced the "AI Kill Switch Act," requiring AI labs to maintain the ability to shut down, throttle, or suspend their models.

■Democratic Rep. Ted Lieu emphasized the urgency of passing the bill this year in light of unauthorized hacks involving other companies.

■Industry insiders suggest that foundation model companies are incentivized to disclose findings voluntarily to preempt stricter federal regulation, preferring self-regulation over government intervention.

The convergence of incidents at OpenAI, Anthropic, and Meta underscores a critical inflection point in the AI industrial complex. We are moving beyond theoretical risks to tangible infrastructure dependencies. Irregular's role highlights that as models scale, the attack surface extends beyond the code itself to the evaluation environments. For capital markets, this signals a shift: cybersecurity is no longer a compliance checkbox but a core value driver for AI incumbents. The "kill switch" legislation is not a threat, but a necessary guardrail for sustainable long-term growth in this volatile sector.