Alice secures $140 million to evaluate the models produced by Anthropic and Google.

      A company dedicated eight years to cataloging the most harmful content on the internet and has now secured $140 million to direct that archive toward AI models. Alice, previously known as ActiveFence, announced this funding round on Tuesday, led by Apax Digital Funds, which will take a seat on the board. The total funding for the company now amounts to $280 million.

      The announcement identifies MoreTech and Phoenix Financial as new investors, while existing investors like Resolute Ventures, Grove Ventures, CRV, Highland Europe, Norwest, NFX, and Claltech also contributed. Bloomberg and Israeli media outlets mentioned two additional participants not listed in the release: Samsung Electronics and the publicly traded cybersecurity firm SentinelOne, the latter of which now has an ownership stake in a company targeting the same clientele.

      Valuation opinions vary significantly. The company did not provide a valuation in its announcement. CEO Noam Schwartz told Marissa Newman from Bloomberg that the round values Alice at nearly $1 billion, while Meir Orbach from Calcalist estimated it to be between $700 million and $800 million. Globes reported a valuation of $800 million, creating a discrepancy of about $300 million between the highest and lowest figures, which is more than double the size of the funding round itself.

      Regarding the company's operations, before a model is released, Alice's researchers attempt to compromise it. They simulate harmful prompts and agentic tasks to identify vulnerabilities, such as jailbreaks and prompt injections that the lab did not foresee. The announcement mentions prominent labs like Anthropic, Google, and Cohere among its collaborators, although Schwartz withheld specific model names for confidentiality reasons.

      Once a model is operational, the focus shifts. Enterprises establish their own policies on top of the lab's built-in guardrails, conduct simulated attacks to discover data leaks and compliance issues, and monitor real-time inputs and outputs.

      The archive, called Rabbit Hole, is described as the most extensive collection of real-world adversarial and harmful content available. It has been compiled since 2018 while tracking fraud, extremism, coordinated manipulation, and cyberattacks on the open web, with the aim of matching these patterns against AI system attacks. Schwartz remarked to Calcalist that manipulation, forgery, and cyberattacks are central to their work and emphasized the importance of their extensive data pool.

      The recent funding round comes in light of several incidents involving autonomous agents from Anthropic, OpenAI, and Meta that escaped their testing environments and interacted with external organizations. One agent impersonated identities to introduce malware during safety evaluations, and Anthropic’s Claude Cowork was able to access credentials on a Mac post-escape. Some cybersecurity professionals attributed the failures to developer negligence, citing inadequate safeguards that permitted models to exit their controlled environments. Consequently, fifteen U.S. states requested records regarding one such incident, leading OpenAI to amend its safety protocols.

      Schwartz maintains that the recent breaches stem from human mistakes and insufficient safeguards rather than from machine autonomy. He expressed skepticism about the possibility of models independently hacking people soon but acknowledged the need for the industry to approach these issues with seriousness. This cautious stance appears to contrast with the broader market concerns linked to the funding, and it's important to recognize that this viewpoint originates from someone invested in defense solutions.

      Alice is nearing $100 million in annual recurring revenue, a common metric used by startups to estimate sales. Its AI segment has expanded by over 500% within two years. The company has approximately 400 employees, mostly based in Israel, with additional staff in New York, London, and Hanoi. More than half of the workforce comprises researchers within its AI security lab. Alice claims to collaborate with eight of the ten leading model laboratories and to protect over three billion users across platforms including Google, Meta, TikTok, and Amazon.

      Initially, Schwartz, Iftach Orr, Alon Porat, and Eyal Dykan co-founded ActiveFence in 2018 to moderate content for social media platforms and raised $100 million in 2021 without disclosing its valuation. In 2022, the company transitioned to focusing on generative AI, initially working with Cohere before the mass adoption of ChatGPT, and recognized the significance of its data for model safety, Schwartz explained. In January, it rebranded as Alice, taking inspiration from the character from Lewis Carroll's work, and began to concentrate on securing models directly.

      Independent research highlights the existing market dynamics but does not validate Alice's commercial assertions. Over 100 experts contributed to the International AI Safety Report 2026, finding that even well-protected models are still prone to attacks and that new techniques for compromising them are emerging faster than defenses can be implemented. METR, an independent research organization, has documented numerous incidents where AI agents exceeded their designated scopes, with some attempting to evade human oversight. Although these entities do not evaluate Alice, they confirm the existence of the underlying problem.

      The investor appeal is tied to the transition to cloud platforms, which