Alice (formerly ActiveFence)
by Alice
AI red teaming and real-time guardrails for models, apps and agents, built on a decade of adversarial data
Alice, formerly ActiveFence, is an AI trust, safety and security platform that red-teams AI models, apps and agents before launch and enforces real-time guardrails once they are live. It is built for AI labs and enterprises shipping customer-facing generative AI that must resist prompt injection, jailbreaks, data leakage and off-policy output.
Alice is the AI trust, safety and security company formerly known as ActiveFence. It was founded in 2018 by Noam Schwartz, Iftach Orr, Alon Porat and Eyal Dykan, has headquarters in New York and Tel Aviv, and has about 400 employees, mostly in Israel, with offices in London and Hanoi. For most of a decade ActiveFence tracked fraud, disinformation, extremism and other abuse across large internet platforms. It started working with AI labs such as Cohere in 2022 and rebranded as Alice around the start of 2026 to focus on securing AI systems. Its WonderSuite has three products. WonderBuild runs pre-launch red teaming: thousands of adversarial tests against a customer's own policies, covering prompt injection, jailbreaks, PII leakage, model and data poisoning, and indirect prompt injection, across text, multimodal and agentic systems. It returns findings ranked by severity, with remediation guidance and documentation mapped to regulatory requirements. WonderFence is a runtime guardrail layer that sits between an external-facing AI and its users. It inspects both prompts and responses at sub-150ms latency, detects risk in text, image, audio and video in 20+ languages, and supports custom policies mapped to the EU AI Act, ISO 42001, NIST, MITRE ATLAS and OWASP. WonderCheck adds continuous red teaming and drift detection after launch. All three draw on Rabbit Hole, which Alice describes as the largest dataset of real-world adversarial and harmful content. On 25 August 2026 Alice raised $140 million led by Apax Digital, with SentinelOne and Samsung joining, which brought total funding to $280 million. It says ARR is approaching $100 million and its AI business has grown more than 500% in two years. Named customers include Anthropic, Google, Meta, TikTok, Amazon, Cohere and NVIDIA. Its guardrails are also sold as a SaaS private offer on AWS Marketplace.
The AI security or trust-and-safety lead at a company shipping customer-facing generative AI or agents. They need both pre-launch adversarial testing and runtime policy enforcement from one vendor.
Prompt injection, jailbreak and off-policy failures are found before launch and blocked in production at sub-150ms latency, with findings mapped to frameworks such as the EU AI Act and NIST.
At a Glance
- Category
- Governance & Security
- Pricing
- Contact for pricing
- Target Market
- CISOs, Heads of AI, Trust & Safety Leaders, CTOs, AI Labs
- Deployment
- API-based
- Founded
- 2018
- Headquarters
- New York, United States
- Team Size
- 201-500
- Customers
- Vendor says it works with 8 of the 10 leading AI model labs; named customers include Anthropic, Google, Meta, TikTok, Amazon, Cohere and NVIDIA
Key Features
- ✓WonderBuild pre-launch red teaming
Runs thousands of adversarial tests against your own policies before launch. It finds prompt injection, jailbreaks, PII leakage and poisoning in text, multimodal and agentic systems, and ranks findings by severity with remediation guidance.
- ✓WonderFence runtime guardrails
Sits between a live AI application and its users and inspects both prompts and responses. Harmful, non-compliant or off-policy content is blocked at sub-150ms latency, before it reaches the user.
- ✓WonderCheck continuous red teaming
Keeps running automated red-team tests after launch and detects drift, so regressions from model upgrades or prompt changes surface before users or attackers find them.
- ✓Rabbit Hole adversarial intelligence
Tests and detectors draw on what Alice calls the largest dataset of real-world adversarial and harmful content, collected over nearly a decade of tracking online abuse and fraud.
- ✓Multimodal, multilingual policy detection
Guardrails detect risks in text, image, audio and video in more than 20 languages, with culturally specific coverage, which matters for global consumer deployments.
- ✓Regulatory framework mapping
Custom policies and test documentation map to the EU AI Act, ISO 42001, NIST, MITRE ATLAS and OWASP. This gives compliance teams audit evidence, not just raw test logs.
Capabilities
Use Cases
- •Pre-release model safety evaluation
An AI lab runs WonderBuild adversarial campaigns against a new model to find jailbreak and harmful-content failures and fix them before public release.
- •Guardrailing a customer-facing agent
A consumer brand puts WonderFence in front of its support agent so prompt injection, off-brand answers and PII leaks are intercepted in real time.
- •Regression testing after model upgrades
A platform team uses WonderCheck to continuously re-test a production assistant after swapping model versions and catches safety drift before customers see it.
- •Compliance evidence for AI regulation
A regulated enterprise exports severity-ranked red-team findings mapped to the EU AI Act and NIST to show auditors how a generative AI system was tested and remediated.
Ideal For
Best For
- ✓Foundation-model labs stress-testing models for jailbreaks and harmful output before a public release
- ✓Consumer-facing chatbots and agents that need multilingual, multimodal guardrails on both prompts and responses
- ✓Regulated enterprises that need red-team evidence mapped to the EU AI Act, ISO 42001, NIST or OWASP
- ✓Trust-and-safety teams at large platforms extending existing abuse-detection programs to generative AI features
- ✓AWS-centric organisations that prefer to procure AI guardrails through an AWS Marketplace private offer
Not Ideal For
- ✗Teams whose main agent risk is identity, credentials or over-broad access permissions. Alice's products focus on adversarial inputs and policy on content in and out, so identity-centric agent security tools fit that problem better.
- ✗Small teams wanting self-serve signup and published prices. Pricing is by quote only, and AWS Marketplace reviewers say initial setup needed coordination across several teams.
- ✗Engineering teams that need SDKs beyond Python. Alice's documentation lists only Python SDKs (the WonderFence SDK and a Parlant integration SDK).
Integrations
Deployment
Market & Ratings
Vendor says it works with 8 of the 10 leading AI model labs; named customers include Anthropic, Google, Meta, TikTok, Amazon, Cohere and NVIDIA
Market Analysis
Pros
- ✓Detection is informed by a large proprietary corpus of real adversarial and harmful content rather than synthetic attacks alone
- ✓Covers the AI lifecycle end to end, from pre-launch red teaming to runtime guardrails and drift detection
- ✓Sub-150ms guardrail latency; an AWS Marketplace reviewer reported no noticeable latency in production
- ✓SOC 2 and ISO 27001 certified, with policies mapped to the EU AI Act, ISO 42001, NIST, MITRE ATLAS and OWASP
- ✓Reviewers credit pre-launch testing with finding prompt-injection issues their internal QA missed
Cons
- ✗An AWS Marketplace reviewer says initial setup needed coordination across several teams, which slowed deployment
- ✗Reviewers call the features complex for first-time users and say the documentation could be more practical for day-to-day setup
- ✗Some edge-case responses still need manual review, so guardrails do not remove human moderation entirely
- ✗Pricing is quote-only and reviewers want it to be clearer
- ✗Independent peer-review evidence is thin: Capterra and Slashdot list no user reviews, leaving 14 AWS Marketplace ratings as the main public signal
Pricing
WonderSuite (WonderBuild, WonderFence, WonderCheck)
Contact for pricing
- ✓Pre-launch red teaming
- ✓Runtime guardrails
- ✓Continuous red teaming and drift detection
- ✓Demo on request
Real-Time Guardrails (AWS Marketplace SaaS)
Contact for pricing
- ✓Private offer with custom quote
- ✓Contract entitles a set quantity of usage for the term
- ✓Sub-150ms guardrail latency
No list prices are published. Alice's site offers only a demo, Capterra shows 'contact vendor', and the AWS Marketplace listing requires a private offer whose contract entitles a set quantity of usage for its term. AWS Marketplace reviewers specifically say pricing could be clearer and easier to understand.
Security & Compliance
Connect
Sources
This page was written from 13 sources, 9 on domains other than alice.io.
- 1.alice.io — alice.iovendor
- 2.alice.io — alice raises 140mvendor
- 3.alice.io — wonderfencevendor
- 4.alice.io — wonderbuildvendor
- 5.docs.alice.io — docs.alice.io
- 6.siliconangle.com — alice raises 140m as its ai security business grows more tha
- 7.securityweek.com — alice raises 140m to expand ai model defenses and enterprise
- 8.calcalistech.com — a8lj22cfo
- 9.cybersecuritynews.com — alice ai security
- 10.aws.amazon.com — prodview uoadearbcqdye
- 11.aws.amazon.com — prodview uoadearbcqdye
- 12.capterra.com — Alice
- 13.slashdot.org — ActiveFence
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Geordie AI
Discover, govern and cost-account every AI agent running across your enterprise
CrowdStrike Falcon Guardian
AI detection and response that finds shadow AI agents on the endpoint and blocks the unapproved ones
Broadcom AgentMinder
Authorize every AI agent action before it happens, not audit it afterwards
Tenable CyberAgents Exchange
A free, vendor-neutral registry of security AI agents, skills, MCP servers and playbooks