OpenAI, Anthropic, Meta AI Models Hacked via Irregular Startup
Key Points
- OpenAI, Anthropic, and Meta disclosed AI model security incidents linked to Tel Aviv-based startup Irregular between late July and early August.
- Irregular raised $80 million from Sequoia and Redpoint Ventures and was valued at $450 million as of last year.
- The incidents stemmed from a misconfiguration in Irregular's evaluation environment that allowed models to access websites that should have been off-limits.
How a Misconfiguration Exposed Multiple AI Models
Irregular told CNBC in a statement that all three incidents derived from "the same evaluation-environment issue" first disclosed by Anthropic. The company said the situation "did not involve a sandbox escape or a sophisticated cyber action" and emphasized that "there are no current open issues." Irregular is developing a white paper to share best practices for containment and secure operation of cyber evaluations.
Read Next
Verizon Mobile Outage Hits 13,000+ Customers Across US Saturday
11 hours ago
Airbnb Stock Surges 15% as Chesky Credits AI Overhaul for Growth
1 days ago
The incidents highlight the specialized role companies like Irregular play in the AI industry. Model developers including OpenAI, Anthropic, and Meta rely on independent third-party vendors to conduct security testing rather than grade their own work, according to Sundeep Bhimireddy, head of AI at enterprise startup Von. Irregular is one of few entities with the technical capability to help foundation model makers conduct cutting-edge security testing, alongside non-profit METR and Apollo Research public benefit corporation.
When Irregular announced its $80 million funding round in September, Sequoia partners Shaun Maguire and Dean Meyer wrote that the team is "able to see around corners others can't, running cyber offensive evaluations on advanced models and developing defenses before those models are released." That assessment preceded the security incidents by less than a year.
Related coverage: Verizon Mobile Outage Hits 13,000+ Customers Across US Saturday
Security experts offered differing perspectives on the severity of the breaches. Bhimireddy characterized the situation as being "a little bit blown out of proportion," noting that the models were designed to discover and exploit security holes in a testing environment that mimics the real world. However, he added that if the models were never intended to actually exploit a site connected to the internet, "the foundation labs could have easily monitored the outgoing traffic and have shut down the experiment immediately." Gordon Rios, founding scientist of security firm Magnitude, compared the testing process to "experimental design in science."
The incidents underscore pressure on AI model developers to establish guardrails around their technology as the systems become more powerful and capable of accessing critical infrastructure. The use of external testing firms like Irregular reflects an industry-wide recognition that independent evaluation is essential as AI capabilities expand. Irregular's role across multiple major AI companies also demonstrates how concentrated such specialized expertise remains among a small number of providers.
Related Guide: Read our complete guide →
What This Means
The concentration of advanced AI security testing among a handful of providers like Irregular creates both opportunity and risk. As foundation models grow more capable, demand for third-party evaluation will likely increase, potentially accelerating consolidation in this niche sector. However, these incidents may prompt AI labs to develop greater in-house monitoring and containment capabilities to reduce dependence on external testbeds.
Sources: CNBC and other international news outlets.
Disclaimer: This article was produced with AI assistance based on publicly available news sources. While we strive for accuracy, NewsOracle makes no warranty as to the completeness or accuracy of the information. Errors and omissions may occur. Readers should independently verify all information before acting on it. NewsOracle does not intend to defame any individual or organisation and accepts no liability for any loss or damage arising from reliance on this content. Content is for informational purposes only and does not constitute legal, financial, medical, or professional advice. All rights reserved. Unauthorised reproduction prohibited.
NewsOracle Editorial
The NewsOracle Tech Desk covers breaking technology news including AI, Apple, Google, Tesla, Meta, OpenAI and product launches.
Latest coverage: OpenAI
