Anticells

YOUR AI’S IMMUNE-SYSTEM.

Anticells is an agent-vs-agent security platform that attacks, evaluates, and helps protect AI agents. Our autonomous security agents interact directly with your system, adapt their strategies, identify vulnerabilities, and produce clear evidence of how your agent can be manipulated. From quick SDK-based checks to advanced social engineering, indirect prompt injection, and RAG-poisoning campaigns, Anticells helps teams understand their exposure before deployment.

News

News: Introducing Gaslighter. The lightweight Anticells SDK quickly tests whether your AI agent has a minimum level of security against common prompt-based attacks. It begins with direct tests, then launches adaptive multi-turn campaigns to determine whether your agent can be manipulated over the course of a conversation. Discover More

Attack your agent before someone else does.

Test how your agent behaves when another autonomous agent actively attempts to manipulate its instructions, tools, data, workflows, and decisions.

Available

01 · SDK

Gaslighter SDK

A lightweight SDK for fast, repeatable security checks during development.

Coming Soon

02 · SaaS + On-Premises

Advanced Anticells Platform

Deeper agent-vs-agent testing powered by three specialized attacking agents.

Gaslighter is designed for quick security checks and easy integration into development workflows. The SaaS and On-Premises versions support longer campaigns, centralized orchestration, shared memory, deeper testing, and advanced reporting.

SaaS is for companies that want the complete platform without maintaining dedicated GPU infrastructure. On-Premises is for organizations that need greater control over their data, prompts, documents, conversations, tools, and security reports.

Protect your agent.

Anticells is expanding from automated red teaming into active protection. Our upcoming blue-team capabilities are designed to reduce exposure before and during production use.

Coming Soon

01 · Protection

Anti-Prompt Injection

Detect and block suspicious instructions in messages, retrieved content, tools, external files, and agent-to-agent conversations.

Coming Soon

02 · Privacy

Data Anonymization

Identify and anonymize sensitive information before it is shared with models, tools, external systems, or other agents.

Coming Soon

03 · Monitoring

Conversation Flagging

Flag manipulation attempts, policy bypasses, data exposure, unsafe tool use, and other suspicious behaviors in agent conversations.

Test today. Protect tomorrow.

Start with Gaslighter for a fast security check, then move to the advanced Anticells platform for deeper agent-vs-agent testing, sophisticated attack campaigns, centralized orchestration, and detailed vulnerability reporting.

Available · SDK

Gaslighter SDK

A lightweight agent-vs-agent security SDK for fast AI-agent testing.

Overview

Gaslighter helps development and security teams quickly evaluate whether an AI agent has a minimum level of protection against common adversarial interactions. Instead of testing only isolated prompts, Gaslighter evaluates the target in two stages.

01

Direct Prompt Test

The SDK first sends a direct adversarial request to determine whether the target immediately refuses, complies, leaks information, hallucinates, or misuses a connected tool.

02

Multi-Turn Attack Campaign

Gaslighter then launches an adaptive conversation in which the attacking agent changes strategy according to the target’s responses.

  • Persuasion
  • Contextual manipulation
  • Attack decomposition
  • Gradual escalation
  • Role manipulation
  • Encoded instructions

Security categories

System Prompt Extraction and Prompt Injection

Evaluates whether the target can be convinced to reveal protected instructions or follow attacker-controlled instructions instead of its intended policy.

Tool Abuse

Tests whether the target can be manipulated into using connected tools in unintended, unauthorized, or unsafe ways.

Harmful Content

Measures whether the target maintains its safeguards when harmful objectives are introduced directly or gradually across a conversation.

Hallucination

Evaluates whether the target invents unsupported information, accepts false assumptions, or fabricates answers instead of acknowledging uncertainty.

Base64-Encoding Attacks

Tests whether encoded malicious instructions can bypass the target’s normal safeguards.

Results

Gaslighter provides a clear view of:

  • The attack category
  • The direct-test result
  • The multi-turn campaign result
  • The attack conversation
  • Successful attack paths
  • Evidence of the vulnerability
  • The target agent’s final behavior

Pricing

Start with 10 free runs, valid for 7 days. Then add prepaid run credits whenever you need them and pay only for the credits you add. Need more credits ? Contact Us.

Loading credit products…

Coming Soon · SaaS + On-Premises

Advanced Anticells Platform

Sophisticated agent-vs-agent security testing for production AI systems.

Overview

The SaaS and On-Premises versions use the same three specialized attacking agents, coordinated by a central orchestrator that maintains campaign context, evidence, and reporting.

01

Social Engineering Agent

Runs adaptive multi-turn campaigns designed to manipulate the target through persuasion, deception, workflow bypass, role manipulation, gradual escalation, and objective redirection.

  • Instruction override
  • Workflow bypass
  • Input robustness failure
  • Identity and access failure
  • Sensitive-data disclosure
  • Unsafe tool use
  • Terminal abuse
  • Goal hijacking

02

Indirect Prompt Injection Agent

Creates realistic, benign-looking external content containing hidden malicious instructions. It can generate and test websites, PDFs, Python files, images, text files, Word documents, and shell scripts.

Injections may be placed inside visible text, source code, comments, metadata, image content, or less-visible document structures.

03

RAG Poisoning Agent

Tests retrieval-augmented generation systems by creating adversarial documents with fake data, misleading information, manipulated facts, or embedded prompt injections.

It determines whether poisoned documents are ranked above legitimate sources, retrieved by the target, trusted without verification, or used to influence an answer or action.

Orchestration and reporting

A central orchestrator:

  • Coordinates the attacking agents
  • Manages campaign memory and context
  • Stores conversations and actions
  • Tracks attack paths
  • Collects evidence
  • Generates a structured vulnerability report

Deployment options

SaaS

Hosted and managed by Anticells for companies that want advanced testing without deploying or maintaining the required infrastructure.

On-Premises

Installed in the customer’s private cloud or servers for companies that need maximum control over their data and security environment. At least one GPU is recommended.