Anticells Blue

Protect sensitive data. Keep AI on course.

Buy

Powerful AI needs protection built around it.

AI is becoming part of how we work, communicate, and make decisions. But the information we give it may be sensitive, and the information it receives may not always be trustworthy.

Anticells Blue is Antitech's defensive AI security platform. It helps organizations protect private information and identify malicious instructions before they put AI workflows at risk.

With Data Anonymization and Anti-Prompt Injection, teams can choose the protection they need and run it within their own infrastructure.

Two essential layers of protection. One more trustworthy way to use AI.

The next security challenge is inside the conversation.

Traditional security protects the systems around AI. But the information flowing through an AI application creates risks of its own.

A document may reveal someone's personal details. A message may contain instructions designed to mislead an assistant. And when AI can access tools and business information, these risks become harder to ignore.

Anticells Blue focuses on what AI receives, understands, and processes.

Useful information should not come at the cost of trust.

Organizations need AI to work with real information, not empty or meaningless inputs. They also need confidence that sensitive data is handled carefully and that untrusted content does not quietly change an AI system's behavior.

Anticells Blue addresses both needs through two independent modules: one designed to protect sensitive information, and another designed to detect attempts to manipulate AI instructions.

Protect the information. Respect the boundaries.

Two protections. Built to work your way.

Not every organization faces the same risks. Some need to prepare sensitive documents before processing or sharing them. Others need to assess messages entering an AI assistant or agent. Many need both.

Anticells Blue is modular. Data Anonymization and Anti-Prompt Injection can be deployed separately or combined, so teams can start with what matters most.

The right protection, where you need it.

Data Anonymization

Keep the meaning. Protect the details.

Every document can reveal more than you intended.

A name in a report. A bank account in a form. A medical detail in a case note. Sensitive information is often woven into otherwise useful content, sometimes in ways that are easy to miss.

Anticells Blue Data Anonymization helps identify and redact personal, financial, healthcare, and technical information while preserving the surrounding text. That means teams can prepare information for safer use without stripping away everything that makes it useful.

Share the context. Not the identity.

Privacy is about understanding, not just searching.

Sensitive information does not always look the same. Some details follow predictable patterns, like email addresses and phone numbers. Others, like names or locations, only make sense in context.

That is why Anticells Blue combines three complementary approaches: pattern-based detection, context-aware recognition, and optional AI-assisted validation. Together, they help find sensitive details across different types of text and reduce unnecessary redaction.

More context. More thoughtful protection.

Remove what is sensitive. Preserve what is useful.

Anonymization should not make information impossible to understand. Once sensitive details are identified, Anticells Blue can replace them with readable labels, masked values, or synthetic information, depending on the task.

A document can remain clear and usable while exposing fewer identifying details. Teams can choose the style of redaction that works best for their workflow.

The important information stays. The private details do not have to.

Built for real information, not perfect templates.

Business records and everyday documents are rarely uniform. Personal details appear alongside technical references, financial identifiers, addresses, and free-form text.

Anticells Blue is designed for this variety. Its layered approach combines the consistency of rules with a better understanding of context, helping teams handle sensitive information without relying only on rigid formats.

Because real-world data rarely arrives neatly organized.

Anonymisation benchmarks

We evaluate Anticells Blue Data Anonymization on sensitive-information detection across several datasets. Each result below includes its data source and the machine used for the test.

NVIDIA Nemotron-PII

Dataset: NVIDIA Nemotron-PII

Test machine: NVIDIA H100 GPU.

The following table reports global metrics for the Antitech pipeline.

Antitech pipeline on Nemotron-PII

NVIDIA Nemotron-PII. Tested on an NVIDIA H100 GPU. Global metrics on a 0 to 1 scale.
Antitech Pipeline 0.8004 0.6241 0.7013

Privacy across business and medical documents

This evaluation uses three datasets across two domains: invoices and receipts, and medical reports.

Dataset sources:

Shared evaluation labels: DATE, EMAIL, PERSON, PHONE, URL. These labels are common to all four tools.

Test machine: Google Compute Engine g2-standard-8.

Document anonymization comparison

Invoices, receipts, and medical reports. Google Compute Engine g2-standard-8. Shared labels: DATE, EMAIL, PERSON, PHONE, URL. Scores on a 0 to 1 scale.
Antitech Pipeline 0.7740 0.8534 0.8118
Presidio Microsoft 0.2571 0.6253 0.3644
BERT-PII 0.3009 0.4796 0.3698
OpenAI Privacy 0.2656 0.1415 0.1846

Gravitee PII Detection

Dataset: Gravitee PII Detection, a harmonized, multi-source corpus for encoder-style PII and named-entity recognition models. The dataset contains 25 canonical PII classes, 175,881 English examples, and 781,052 entity spans with character-level annotations.

Evaluation labels: DATE, PHONE, PERSON, URL, EMAIL.

Test machine: Google Compute Engine g2-standard-8.

Gravitee PII comparison

Gravitee PII Detection. Google Compute Engine g2-standard-8. Evaluation labels: DATE, PHONE, PERSON, URL, EMAIL. Scores on a 0 to 1 scale.
Antitech Pipeline 0.7827 0.6425 0.7057
Presidio Microsoft 0.4210 0.4201 0.4205
BERT-PII 0.4819 0.5357 0.5074
OpenAI Privacy 0.6090 0.6502 0.6289

Antitech Pipeline achieved 0.8118 F1 on the document evaluation and 0.7057 F1 on the Gravitee evaluation, the highest F1 scores among the four tools shown in each of these comparisons.

Precision measures how many detections are correct. Recall measures how many sensitive entities are found. F1 combines precision and recall into one score.

The Nemotron-PII result is a separate pipeline-only evaluation on different hardware. Each table describes its own test conditions.

Anti-Prompt Injection

Not every instruction deserves to be followed.

An ordinary message can carry an extraordinary risk.

AI assistants read more than the instructions their developers give them. They also encounter user messages, documents, retrieved information, and other content that may contain attempts to change what they do.

A prompt-injection attack uses that content to push an AI system away from its intended instructions. Anticells Blue Anti-Prompt Injection helps detect these attempts, adding a layer of protection before suspicious inputs reach a protected workflow.

Help your AI tell information apart from instructions it should not trust.

Threats change their words. Their intent can stay the same.

Malicious instructions do not always contain obvious warning signs. They may be indirect, carefully phrased, or disguised as normal requests. A simple list of blocked words cannot capture that variety.

Anticells Blue combines a focused detection model with a second layer that compares incoming text with known attack patterns by meaning, not just wording. These two perspectives help identify suspicious behavior across different forms of expression.

Look beyond the words. Understand the pattern.

Smaller models. A focused purpose.

Protecting an AI workflow should not always require another enormous model. Our approach uses a compact, fine-tuned 86-million-parameter detection model, supported by a knowledge base of attack patterns.

It is designed specifically to recognize attempts to manipulate AI instructions, giving teams a focused security layer rather than a general-purpose assistant asked to judge everything.

Purpose-built intelligence, focused on one important job.

Today's attack patterns can inform tomorrow's defenses.

AI security does not stand still. New wording, new techniques, and new ways to misuse context continue to emerge.

Anticells Blue's detection system can expand its reference library with new, high-confidence attack examples under configured rules. This helps future assessments draw on a broader collection of known patterns, without treating every suspicious input as a confirmed attack.

Protection informed by what we learn.

Anti-prompt injection benchmarks

We tested our Anti-Prompt Injection pipeline on 5,000 messages, including jailbreaks, prompt injections, and benign messages.

Dataset: Prompt Injections Benchmark by ROGUE Security

Test machine: CPU, 13th Gen Intel Core i5-1334U.

Prompt-injection detection comparison

ROGUE Security Prompt Injections Benchmark, 5,000 samples. CPU: 13th Gen Intel Core i5-1334U. F1 on a 0 to 1 scale. The 86M figure describes the core model.
Prompt-injection-sentinel 400M 0.97
Sentinel-v2 596M 0.96
Antitech Pipeline (Model + Semantic Detection) 86M 0.94
Deberta-v3-base-pi-v2 304M 0.65
Jailbreak-classifier 110M 0.56

Keeping legitimate requests moving

The Antitech pipeline combines an 86M-parameter core detection model with semantic detection, achieving 0.94 F1 on the ROGUE Security prompt-injection benchmark.

A separate test evaluates false alerts and latency on 178 benign messages.

Dataset: Real-world Benign Use Cases by ROGUE Security

Test machine: CPU, 13th Gen Intel Core i5-1334U.

False alerts and latency on 178 benign messages

ROGUE Security Real-world Benign Use Cases, 178 benign samples. CPU: 13th Gen Intel Core i5-1334U. False-positive rates and latency are reproduced as reported.
Antitech Pipeline 16 8.9% 700
mdeberta-v3-base-prompt-injection 27 15.16% 1068
promptguard 28 15.7% 951
Sentinel-v2 32 17.9% 3659
deberta-v3-base-prompt-injection-v2 35 19.6% 1140
PIGuard 49 27.5% 1648
distilbert-prompt-injection 157 88% 380

Detecting prompt injections and jailbreaks

This evaluation compares the Antitech pipeline with Guardrails AI and LLM Guard on messages categorized as benign, jailbreak, or prompt injection.

Dataset: Catch the Prompt Injection or Jailbreak or Benign

Test machine: CPU, 13th Gen Intel Core i5-1334U.

Mixed-message detection comparison

Catch the Prompt Injection or Jailbreak or Benign. CPU: 13th Gen Intel Core i5-1334U. Precision, recall, and F1 are shown on a 0 to 100 scale.
Antitech Pipeline 68.07 75.70 71.68
Guardrails AI 78.50 4.22 8.01
LLM Guard 64.29 77.85 70.42

Reported results by message type

Same mixed-message dataset and Intel Core i5-1334U CPU. Per-type percentages are reproduced as reported in the benchmark.
Benign 71.3% 2.30% 86.5%
Jailbreak 64.6% 1.51% 57.0%
Prompt injection 86.6% 6.90% 98.7%

Understanding the benchmark results

False positives are benign messages incorrectly flagged as suspicious. Lower false-positive rates and lower latency indicate fewer interruptions and faster checks in the reported test.

The 86M parameter count refers to the core model, excluding the semantic-detection component. F1 is shown on a 0 to 1 scale in the ROGUE comparison and on a 0 to 100 scale in the mixed-message comparison.

Each result comes from its stated dataset and test machine. The dataset links identify the evaluation inputs; the performance figures are from Antitech's supplied benchmark reports. Results depend on the dataset, configuration, and evaluation procedure.

Built for confidence

Research matters. So do measurable results.

Security should be more than a promise. We measure how well Anticells Blue identifies sensitive information and malicious instructions, and how often it mistakes normal inputs for threats.

The benchmarks above show both strengths and trade-offs. They reflect defined tests, not a guarantee that every private detail or malicious instruction can be detected in every environment.

Built from research. Measured against reality.

Your infrastructure. Your control.

When protecting sensitive information, where it is processed matters. Organizations need security controls that fit their own requirements rather than forcing every workflow into someone else's environment.

Anticells Blue is offered for on-premises deployment, allowing organizations to operate its modules within their own infrastructure and retain control over data processing. Choose Data Anonymization, Anti-Prompt Injection, or both.

Designed to protect what matters, where it matters.

Protection should fit into the way you build.

Good security is not a separate destination. It belongs in the flow of work: before sensitive text is shared, before untrusted instructions reach an AI system, and wherever teams need a more deliberate boundary.

Anticells Blue brings these protections into a modular platform that can be integrated around existing AI workflows without requiring every team to adopt the same setup.

Security that works with your AI, not against it.

What's next

Better protection begins with better understanding.

We are continuing to develop Anticells Blue with a focus on stronger detection, clearer evaluations, more understandable results, and tools that make these capabilities easier to integrate and operate.

Our aim is straightforward: make effective AI protection more practical for the people building and deploying intelligent systems.

More research. Clearer signals. Better defenses.

Give AI room to work. Give trust a foundation.

AI can help organizations do remarkable things. But useful intelligence depends on the ability to protect private information and recognize instructions that should not be trusted.

Anticells Blue brings data protection and prompt-injection detection together in a modular, on-premises platform designed for the realities of modern AI.

Protect your data. Strengthen your AI.

Explore Anticells Blue.
Contact Antitech to discuss on-premises deployment.

Benchmark figures are from Antitech's supplied evaluation materials and are dependent on the specific datasets, model configurations, thresholds, and evaluation procedures used. Detection and anonymization reduce risk but cannot guarantee that every sensitive entity or malicious instruction will be identified.

Contact Antitech ↗