> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google and Anthropic Just Paid $140M Because They Can't Secure Their Own AI
- URL: https://wire.fourthweb.ai/google-and-anthropic-just-paid-140m-because-they-cant-secure-their-own-ai/
- Published: 2026-08-25T13:00:00.000Z
- Updated: 2026-08-25T13:30:46.000Z
- Description: The companies building the most powerful AI models just admitted they can't secure them alone. Alice, an Israeli AI safety startup working with Google and Anthropic, raised $140 million following recent high-profile AI system hacks
- Author: Travis Wright
- Tags: AI Agent Economy, Agentic Workflows, AI Agents, AI Infrastructure, AI Governance, OpenAI, Anthropic, IPO Watch, Funding Rounds

**The companies building the most powerful AI models just admitted they can't secure them alone.**

### The Summary

- [Alice, an Israeli AI safety startup working with Google and Anthropic, raised $140 million](https://www.bloomberg.com/news/articles/2026-08-25/ai-safety-startup-alice-partner-of-google-anthropic-raises-140-million?ref=wire.fourthweb.ai) following recent high-profile AI system hacks
- The raise signals that model security is now a specialized vertical, not just an internal team problem
- Big tech's willingness to outsource AI safety suggests the threat surface is growing faster than their ability to defend it

### The Signal

[Alice's $140 million raise](https://www.bloomberg.com/news/articles/2026-08-25/ai-safety-startup-alice-partner-of-google-anthropic-raises-140-million?ref=wire.fourthweb.ai) comes at an inflection point. The timing matters. This isn't precautionary funding for theoretical risks. This follows actual breaches, actual exploits, actual moments when AI systems did things their builders didn't intend.

The partnerships tell the story. Google and [Anthropic](https://wire.fourthweb.ai/tag/anthropic/) aren't licensing this tech for compliance theater. They're integrating third-party security into their core model infrastructure because the attack vectors have multiplied beyond what internal red teams can cover. Prompt injection, model extraction, jailbreaks that bypass guardrails. The threat landscape evolved from academic papers to weaponized toolkits in under 18 months.

> "When the companies building frontier models start buying security from specialists, the threat is no longer hypothetical."

Here's what changed: AI models are now valuable enough to hack and accessible enough to exploit. A GPT-4 class model represents billions in training costs and strategic advantage. Competitors want to steal the weights. Bad actors want to remove the safety filters. Nation states want both. The surface area is massive: API endpoints, fine-tuning pipelines, the sprawling supply chain of datasets and [compute](https://wire.fourthweb.ai/tag/ai-infrastructure/).

Alice is building for a world where models are infrastructure, not products. Infrastructure needs security specialists. You don't build your own firewall vendor. You don't roll your own encryption. And increasingly, you don't defend your own frontier AI models with just internal teams.

**Key differences from traditional cybersecurity:**

- AI systems can be attacked through natural language, not just code exploits
- Model behavior degrades subtly under poisoning attacks, hard to detect
- Fine-tuning creates thousands of derivative models, each a potential vulnerability

The $140 million also reflects where we are in the agent economy buildout. As models get embedded into actual workflows, into [autonomous agents](https://wire.fourthweb.ai/tag/ai-agents/) handling real transactions and sensitive data, the security requirements compound. An agent that can book travel or execute trades needs hardened infrastructure. The liability exposure is different. The blast radius of a compromised agent is wider than a compromised database.

This raise is validation that AI safety is a business category, not a research lab. The money will go toward detection systems, automated red teaming, continuous monitoring of model outputs for drift or manipulation. These are products now, with revenue models and enterprise contracts.

### The Implication

Watch for AI safety to bifurcate into two camps: the researchers focused on alignment and existential risk, and the security vendors focused on near-term exploits and enterprise protection. Alice is firmly in the second camp, and that's where the immediate commercial demand lives.

For companies building on frontier models, this changes the security budget conversation. Model access isn't a fixed cost anymore. It's model access plus monitoring plus red teaming plus continuous defense. Factor that into your agent economics before you scale.

### Sources

[Bloomberg Tech](https://www.bloomberg.com/news/articles/2026-08-25/ai-safety-startup-alice-partner-of-google-anthropic-raises-140-million?ref=wire.fourthweb.ai)