Alice raised $140m to red-team the frontier, and a security vendor bought in quietly

The pre-ship half of this business is where the structural consequence sits. Alice's researchers work a model over before it ships, simulating malicious prompts and agentic tasks and hunting for jailbreaks, prompt injection and behaviour the lab did not intend ^[6]. The company names Anthropic, Google and Cohere among the labs it works with, and its chief executive declined to tell Bloomberg which specific models, citing confidentiality ^[7].

The product being bought is an archive. Alice calls it Rabbit Hole and describes it as the largest body of real-world adversarial and harmful content anywhere, assembled by tracking fraud, extremism, coordinated manipulation and cyberattacks across the open web since 2018 ^[23]. Noam Schwartz's own framing is that there are infinite ways to break a model and you cannot defend against something you have never seen ^[9]. Take that seriously from the buyer's side and the coverage of any pre-ship evaluation is bounded by what the collection happens to hold. In-house red teams miss different things at different companies. A shared supplier misses the same things everywhere it is installed.

The pricing is soft in a way the round's size hides. Against revenue approaching $100mn a year, growing more than 500% on the AI side in two years ^[24], the reported valuations put the business somewhere above 7x sales at the low end and above 10x at the chief executive's own figure ^[26]. Half of every dollar Alice has raised arrived in this one round ^[19], and the headcount tells the same story: more than 150 researchers inside a company of about 400, most of them in Israel ^[15].

The investor list is where the disclosure thins out. The release names MoreTech and Phoenix Financial as the new participants, alongside existing backers including Resolute Ventures, Grove Ventures, CRV and Norwest ^[3]. SentinelOne, a listed cybersecurity company, now holds a position in a firm whose post-launch business is selling enterprises policy enforcement on top of lab guardrails, simulated attacks against their own deployments, and live monitoring of inputs and outputs ^[8]. That is a security budget line, sold to security buyers, with a public security vendor on the cap table. It reached the market through reporters rather than through either company.

Schwartz is not selling catastrophe. He attributes the recent episodes to human error and inadequate guardrails rather than machine autonomy, and told Bloomberg he does not expect to see models running around hacking people anytime soon, while arguing the industry must take the problem seriously ^[14]. Cybersecurity experts have made a compatible argument, blaming developers whose sandboxes were loose enough to let test models out ^[12]. The pressure that actually moves budgets is procedural: fifteen US states demanded records of one incident, and OpenAI rewrote its safety rules afterwards ^[13]. Demands for records are answered with documentation and a vendor name, which is the market Alice sells into. The archive itself was built to moderate social platforms when the company was ActiveFence, founded in 2018 ^[16], repointed at generative models in 2022 starting with Cohere ^[17], and renamed in January ^[18].